Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Comparative persistence of tumor necrosis factor alpha antagonists in patients with rheumatoid arthritis Fisher, Anat 2012-12-31

You don't seem to have a PDF reader installed, try download the pdf

Item Metadata

Download

Media
ubc_2012_fall_fisher_anat.pdf [ 2.61MB ]
Metadata
JSON: 1.0073059.json
JSON-LD: 1.0073059+ld.json
RDF/XML (Pretty): 1.0073059.xml
RDF/JSON: 1.0073059+rdf.json
Turtle: 1.0073059+rdf-turtle.txt
N-Triples: 1.0073059+rdf-ntriples.txt
Original Record: 1.0073059 +original-record.json
Full Text
1.0073059.txt
Citation
1.0073059.ris

Full Text

COMPARATIVE PERSISTENCE OF TUMOUR NECROSIS FACTOR ALPHA ANTAGONISTS IN PATIENTS WITH RHEUMATOID ARTHRITIS   by Anat Fisher M.D. The Hebrew University Jerusalem Israel, 1990   A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY  in  THE FACULTY OF GRADUATE STUDIES (Pharmacology and Therapeutics)   THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  August 2012 ©Anat Fisher, 2012    ii ABSTRACT Introduction: This thesis comprises three studies examining treatment persistence with tumour necrosis factor alpha (TNFα) antagonists in rheumatoid arthritis (RA) patients. Persistence, also commonly known as duration of treatment, has been suggested as an indirect measure that balances benefit  and harm, and it is important for cost-effectiveness analysis and budget planning. Previous research has examined the effect of drug and patient characteristics on persistence with TNFα antagonists. Objectives: 1) To estimate pairwise comparative persistence with TNFα antagonists infliximab, adalimumab and etanercept in RA patients; 2) To evaluate the impact of investigator factors (methods) and of prescriber factors (propensity for discontinuation and preference for prescribed drug) on comparative persistence estimates with TNFα antagonists in RA patients . Methods: To address these objectives, three population-based studies of an RA cohort were conducted in British Columbia patients using survival analysis methodology. Results: 1) In RA patients similar persistence was observed with infliximab, adalimumab and etanercept, with median persistence equaling approximately 3.5 years; 2) The length of ‘drug-free interval’ used to ascertain drug discontinuation influences the observed magnitude and significance of comparative persistence estimates; 3) Physician factors (prescribing habits) are predictors and possible confounding factors in studies of persistence and comparative persistence.     iii Conclusions: This thesis provides evidence showing that researchers and physicians can influence estimated and actual comparative persistence on TNFα antagonists in RA patients. The role of researchers and physicians in affecting estimates of persistence may explain heterogeneity in these estimates across different studies. Improved methodology in conducting comparative persistence research is needed to establish a high quality body of evidence for the use of patients, clinicians, researchers and policy makers.     iv PREFACE Sections of this thesis are multi-authored manuscripts intended for publication in peer- reviewed journals. Details of authors’ contributions for Chapters 1 to 6 are as follows.  Chapter 1: Anat Fisher was responsible for design, literature search, collation and summary of papers, retrieval of articles, review of studies, and writing of Chapter 1. Colin Dormuth, James Wright and Ken Bassett provided critical review of Chapter 1. Chapter 2: Anat Fisher was responsible for design, literature search, collation and summary of papers, retrieval of articles, review of studies, and writing of Chapter 2. Greg Carney contributed information to the section on the British Colimbia Ministry of Health Administrative Databases. Colin Dormuth provided critical review of Chapter 2. Chapter 3: Anat Fisher was responsible for study concept and design, data management and statistical analyses, interpretation of results, writing Chapter 3, and revisions. Ken Bassett contributed to study concept, formation of research question, and critical review of Chapter 3. Colin Dormuth provided critical review of Chapter 3. Chapter 4: Anat Fisher was responsible for study concept and design, data management and statistical analyses, interpretation of results, writing Chapter 4, and revisions. Colin Dormuth, James Wright, and Ken Bassett provided critical review of Chapter 4. Chapter 5: Anat Fisher was responsible for study concept and design, data management and statistical analyses, interpretation of results, writing Chapter 5, and revisions. Colin Dormuth, James Wright, and Ken Bassett provided critical review of Chapter 5.    v Chapter 6: Anat Fisher was responsible for integration of research results and formation of conclusions from aforementioned results. Colin Dormuth, James Wright, and Ken Bassett provided critical review of Chapter 6.  The research was approved by The University of British Columbia Clinical Research Ethics Board; UBC CREB number H11-00303.    vi TABLE OF CONTENTS Abstract .............................................................................................................................. ii Preface ................................................................................................................................ iv Table of Contents .............................................................................................................. vi List of Tables ..................................................................................................................... ix List of Figures ..................................................................................................................... x List of Abbreviations ........................................................................................................ xi Glossary ............................................................................................................................ xv Acknowledgements ....................................................................................................... xxii Dedication ...................................................................................................................... xxiv Chapter 1 :  Introduction ...................................................................................................... 1 1.1 Conceptual Framework and Previous Research ..................................................... 4 1.2 Thesis Goals ............................................................................................................. 9 1.3 Overview of Thesis Chapters ................................................................................. 10 Chapter 2 :  Methodological Issues in Estimating Persistence .......................................... 12 2.1 General Methodological Background.................................................................... 12 2.1.1 Terminology .......................................................................................................... 13 2.1.2 Measuring Drug Persistence .................................................................................. 16 2.1.3 Comparative Analysis of Persistence .................................................................... 28 2.2 A Methodological Frame for Estimating Comparative Persistence Using British Columbian Ministry of Health Databases ............................................................. 30 2.2.1 Data Source: The British Columbian Ministry of Health Databases ..................... 31 2.2.2 Creating the Study Cohort ..................................................................................... 35 2.2.3 The Exposure and Covariates ................................................................................ 39 2.2.4 Outcome Variables ................................................................................................ 47 2.2.5 Analytic Approach ................................................................................................. 54 Chapter 3 :  Estimating Comparative Persistence of TNFα Antagonists – A Population-  Based Study .............................................................................................................. 56 3.1 Introduction ........................................................................................................... 56 3.2 Objectives and Hypothesis ..................................................................................... 57 3.3 Patients and Methods............................................................................................. 57 3.3.1 Data Source and Study Cohort............................................................................... 57 3.3.2 Exposure and Covariates ....................................................................................... 59 3.3.1 Outcomes ............................................................................................................... 60 3.3.2 Statistical Methods................................................................................................. 62 3.4 Results .................................................................................................................... 64 3.4.1 Comparative Persistence With TNFα Antagonists ................................................ 64 3.4.2 Predictors of Drug Discontinuation ....................................................................... 68 3.4.3 Confounding .......................................................................................................... 71    vii 3.5 Discussion .............................................................................................................. 72 3.5.1 Similar Persistence With Infliximab, Adalimumab and Etanercept ...................... 72 3.5.2 Possible Confounders ............................................................................................ 75 3.5.3 Strengths and Weaknesses of the Study ................................................................ 77 3.6 Conclusions ............................................................................................................ 85 Chapter 4 :  Sensitivity of Comparative Persistence Estimates to the Methods Used to Ascertain Discontinuation .............................................................................. 86 4.1 Background ............................................................................................................ 86 4.2 Study Goals ............................................................................................................ 87 4.3 Patients and Methods............................................................................................. 88 4.3.1 Data Source and Study Cohort............................................................................... 88 4.3.2 Exposure and Covariates ....................................................................................... 89 4.3.3 Outcomes ............................................................................................................... 91 4.3.4 Sensitivity Analysis ............................................................................................... 92 4.3.5 Statistical Methods................................................................................................. 92 4.3.6 Analysis of Reinitiation ......................................................................................... 93 4.4 Results .................................................................................................................... 96 4.4.1 The Study Cohort and Baseline Characteristics ..................................................... 96 4.4.2 Effect of Increased Length of Drug-Free Interval ................................................. 96 4.4.3 Analysis of Reinitiation ....................................................................................... 105 4.4.4 Effect of Covariates ............................................................................................. 108 4.5 Discussion ............................................................................................................ 108 4.5.1 Ascertainment of Drug Discontinuation .............................................................. 109 4.5.2 Possible Superiority of Infliximab ....................................................................... 114 4.5.3 Persistence as a Measure That Integrates Benfit and Harm ................................. 117 4.5.4 Strengths and Limitations of the Study ................................................................ 122 4.5.5 Clinical and Health Policy Implications .............................................................. 122 4.5.6 Research Implications and Directions ................................................................. 123 4.6 Conclusions .......................................................................................................... 125 Chapter 5 :  Prescriber Preference for Tumour Necrosis Factor Antagonists Predicts Persistence on Treatemnt in Rheumatoid Arthritis ...................................... 128 5.1 Background .......................................................................................................... 128 5.2 Objectives and Hypothesis ................................................................................... 129 5.3 Patients and Methods........................................................................................... 130 5.3.1 Data Source and Study Cohort............................................................................. 130 5.3.2 Exposure and Covariates ..................................................................................... 132 5.3.3 Outcome............................................................................................................... 136 5.3.4 Statistical Analysis............................................................................................... 137 5.3.5 Analysis of Switchers .......................................................................................... 139 5.4 Results .................................................................................................................. 140 5.4.1 Study Cohort and Baseline Characteristics .......................................................... 140 5.4.2 Effect of Prescriber Preference on Persistence .................................................... 141 5.4.3 Effect of Prescriber Preference on Comparative Persistence ............................... 146 5.4.1 Analysis of Switchers .......................................................................................... 147 5.5 Discussion ............................................................................................................ 148 5.5.1 Prescribing Habits and Prescribing decisions ...................................................... 149 5.5.1.1 Treatment Initiation ........................................................................................ 149 5.5.1.2 Treatment Adjustment .................................................................................... 153 5.5.1.3 Drug Discontinuation ..................................................................................... 153 5.5.2 Effect of Prescribing Preference .......................................................................... 156    viii 5.5.3 Effect of Prescriber Preference for the Prescribed Drug on TNFα Antagonists            Persistence ........................................................................................................... 160 5.5.4 Effect of Prescriber Preference on Comparative Persistence with TNFα          Antagonists .......................................................................................................... 163 5.5.5 Analysis of Switchers .......................................................................................... 164 5.5.6 Strengths and Limitations .................................................................................... 165 5.6 Conclusions .......................................................................................................... 166 Chapter 6 :  Concluding Chapter ..................................................................................... 167 6.1 There Were Similar Persistence Levels Observed With Infliximab, Adalimumab and Etanercept but Improved Compliance Only With Infliximab in RA Patients 167 6.2 Comparative Persistence with TNFα Antagonists was Sensitive to the Method Used to Ascertain Drug Discontinuation ............................................................. 170 6.3 Physician Prescribing Preferences are Predictors and Possible Confounders in Comparative Persistence Studies ......................................................................... 171 6.4 Confounding is a Threat to Validity in Estimating Comparative Persistence ..... 172 6.5 Interpretation of Persistence and Comparative Persistence as Measures of Benfit-Harm Balance or Relative Benfit-Harm Balance With TNFα Antagonists in RA Patients ...................................................................................................... 175 6.6 Relevance to Studies of Therapeutic Outcomes ................................................... 177 Bibliography ................................................................................................................... 179 Appendices ...................................................................................................................... 214 Appendix A :  Clinical and Pharmacological Background (Supporting Material) ..... 215 Appendix B :  Chapter 3 (Supporting Material).......................................................... 270 Appendix C :  Chapter 4 (Supporting Material).......................................................... 275 Appendix D :  Chapter 5 (Supporting Material) ......................................................... 281     ix LIST OF TABLES Table 1: Three approaches to treatment persistence (adapted from Halpern 2006) ......................... 24 Table 2: Published RA disease algorithms ....................................................................................... 36 Table 3: The rationale for variables included in the regression model ............................................ 43 Table 4: List of covariates included in the final models .................................................................. 61 Table 5: Baseline characteristics ...................................................................................................... 65 Table 6: Hazard ratios for discontinuing TNFα antagonists ............................................................ 68 Table 7: Statistical significance of covariates .................................................................................. 69 Table 8: Confounding effects ........................................................................................................... 71 Table 9: Kaplan Meier estimates of persistence for increasing drug-free intervals ....................... 102 Table 10: Hazard ratios for increasing drug-free intervals ............................................................ 104 Table 11: Analysis of patients reinitiating the same drug (treatment initiated 2001-2008) ........... 106 Table 12: Analysis of patients reinitiating the same drug (treatment initiated 2005-2008) ........... 107 Table 13: List of covariates included in the final models .............................................................. 134 Table 14: Baseline characteristics .................................................................................................. 142 Table 15: Hazard ratios for drug discontinuation, higher PPD versus lower PPD ........................ 145 Table 16: Possible confounder effect of PPD on comparative persistence .................................... 147 Table 17: Analysis of switchers ..................................................................................................... 148 Table 18: Antirheumatic drugs in Canada...................................................................................... 219 Table 19: Pharmacological characteristic of the individual TNFα antagonists ............................. 237 Table 20: Pharmacokinetics characteristics of individual TNFα antagonists ................................ 239 Table 21: Indirect comparisons of efficacy and safety of the individual TNFα antagonists in RA 243 Table 22: Summary of clinical guidelines - indications for discontinuing TNFα antagonists ....... 253 Table 23: Reasons for discontinuing TNFα antagonists ................................................................ 256 Table 24: Discontinuing TNFα antagonists by reason ................................................................... 260 Table 25: Dose escalation .............................................................................................................. 265 Table 26: Dose escalation (additional Studies, from 2006) ........................................................... 266 Table 27: Dose reduction ............................................................................................................... 268 Table 28: List of ICD-9 and ICD-10 codes used in the study ........................................................ 271 Table 29: Sensitivity analysis in studies of TNFα antagonists in rheumatoid arthritis .................. 280    x LIST OF FIGURES Figure 1: Treatment persistence within the framework of pharmacoepidemiological research......... 3 Figure 2: Factors that influence treatment persistence (conceptual framework) ............................... 6 Figure 3: Factors that influence comparative persistence (conceptual framework ) .......................... 7 Figure 4: Persistence, adherence and compliance ............................................................................ 14 Figure 5: Establishing discontinuation based on days-supply and drug-free interval ...................... 22 Figure 6: Algorithm used to establish the number of days-supply for patients treated with       infliximab ....................................................................................................................... 49 Figure 7: Algorithm used to establish the number of days-supply for patients treated with    adalimumab .................................................................................................................... 51 Figure 8: Algorithm used to establish the number of days-supply for patients treated with       etanercept ....................................................................................................................... 52 Figure 9: Calculating days-supply in dose escalation may result in persistence overestimation ..... 53 Figure 10: Drug persistence curves .................................................................................................. 67 Figure 11: Calculating "added persistence" ..................................................................................... 95 Figure 12: Persistence by algorithm to ascertain discontinuation .................................................... 98 Figure 13: Calculating the proportion used to assign the value of PPD ........................................ 133 Figure 14: The product limit estimates of persistence ................................................................... 143 Figure 15: Adjusting for physician preference in studies of treatment initiation ........................... 157 Figure 16: Adjusting for physician preference in studies of drug discontinuation ........................ 158 Figure 17: Cascade of processes from administration of TNFα antagonists to clinical response .. 227 Figure 18: The structure of an antibody ......................................................................................... 230 Figure 19: The structure of the TNFα antagonists ......................................................................... 231 Figure 20: Steady-state concentration-time profiles of infliximab, adalimumab and etanercept... 232 Figure 21: Binding of TNFα antagonists to TNFα ......................................................................... 234 Figure 22: Forest plot for the risk of discontinuing TNFα antagonists due to inefficacy and            safety (adverse event) .................................................................................................. 259 Figure 23: Steps in identifying the study cohort ............................................................................ 270 Figure 24: Assigning values for the outcome variables (five scenarios) ....................................... 271 Figure 25: Persistence by algorithm to ascertain discontinuation  (index date 2005-2008) .......... 276 Figure 26: Persistence with the first course of TNFα antagonists in rheumatoid arthritis patients,         by PPD level ................................................................................................................ 282     xi LIST OF ABBREVIATIONS  $: Canadian dollars G: microgram(s) ACR: American College of Rheumatology ACR20: American College of Rheumatology 20% improvement ACR50: American College of Rheumatology 50% improvement ACR70: American College of Rheumatology 70% improvement ANOVA: analysis of variance ATTRACT (page 248): Anti-TNF Trial in Rheumatoid Arthritis with Concomitant Therapy AUC (Table 20, page 239): area under the curve AUC0-72 H (Table 20, page 239): area under the curve between 0-72 hours after administration BC: British Columbia BIW (Table 20, page 239): twice a week CCI (page 33): Correct Coding Initiative CCP (page 33): Canadian Classification of Diagnostic, Therapeutic and Surgical  CI: confidence interval CIHI (page 33): Canadian Institute of Health Information CMAX (Table 20, page 239): peak concentration CORRONA: Consortium of Rheumatology Researchers of North America COX-2 (page 151): cyclooxygenase-2  CRP: C-reactive protein  DAD: Discharge Abstract Databases DAS: Disease Activity Score DAS28: Disease Activity Score in 28 joints     xii DMARD: disease modifying antirheumatic drug DNA: deoxyribonucleic acid  EOW (Table 20, page 239): every other week ESR (page 242): erythrocyte sedimentation rate  EULAR: European League Against Rheumatism FAB (page 230): Fragment, antigen binding FC (page 229): Fragment, crystallizable GM-CSF (page 226): Granulocyte-macrophage colony-stimulating factor H: hour(s) HACA (Table 19, page 237): human antichimeric antibodies  HAQ: Health Assessment Questionnaire HAQ-DI (Table 22, page 253): Health Assessment Questionnaire Disability Index HLA (page 226): human leukocyte antigen ICBC (page 32): Insurance Corporation of British Columbia ICD: International Statistical Classification of Diseases and Related Health Problems ICD-10: International Statistical Classification of Diseases and Related Health Problems, version 10 ICD-10-CA (p. 271): International Statistical Classification of Diseases and Related Health Problems, 10th Revision, Canadian version ICD-9: International Statistical Classification of Diseases and Related Health Problems, version 9 IGG1 (page 229): immunoglobulin G type 1 IQR (page 17):  interquartile ranges ITT (page 118): intention to treat  KD (Table 19, page 237): antigen dissociation constant KG: kilogram(s)    xiii L: liter(s) LOE (Table 23, page 256): loss of efficacy MD (Table 21, page 243) - means change scores from baseline of mTSS between active and control groups MG: milligram(s) ML: milliliter(s) MSP: Medical Service Plan MTSS (Table 21, page 243): modifications of Total Sharp scores MTX: methotrexate N: number of participants N/A: not available N/R: not reported NF-ΚB (page 225): nuclear factor kappa B NG: nanograms(s)  NR (Table 23, page 256): no response NSAID: nonsteroidal anti-inflammatory drug NSS: not statistically significant  P55 (page 225) : tumour necrosis factor alpha receptor type 1 P75 (page 225): tumour necrosis factor alpha receptor type 2 PPD: prescriber preference for the prescribed drug PPV (page 35): positive predictive value  RA: rheumatoid arthritis RABBIT (page 248): Rheumatoid Arthritis – Observation of Biologic Therapy (in German) RCT: randomized clinical trial RNA (page 220): ribonucleic acid    xiv SAE (Table 21, page 243): serious adverse events SAS: Statistical Analysis System SD: standard deviation  SDAI (Table 22, page 253): Simple Disease Activity Index TACE: tumour necrosis factor alpha converting enzyme TMAX: time to peak concentration TNFα: tumour necrosis factor alpha TNF-R1: tumour necrosis factor alpha receptor type 1 TNF-R2: tumour necrosis factor alpha receptor type 2 TRAF (page 225): tumour necrosis factor alpha receptor associated factor U.S.: United States (of America) VAS: visual analogue scale VS.: versus WDAE: withdrawals due to adverse events WHO: World Health Organization     xv GLOSSARY  ACCURACY (of measurement): “The degree to which a measurement or an estimate based on measurements represents the true value of the attribute that is being measured” [1]. ADDITIVE RISK MODEL: A model in which the combined effects on the risks or rates of several factors is the sum of the effects that would be produced by each of these factors in the absence of others [1].    ADHERENCE: “ Health-related behavior that adheres to the recommendations of a doctor, other health care provider, or investigator in a research project. The word adherence aims to avoid the authoritarian associations of compliance, formerly used to describe this behavior” [1]. In the context of this thesis, adherence describes the overall behavior, in terms of both compliance and persistence.  ALGORITHM: “A systematic process that consists of an ordered sequence of steps with each step depending on the outcome of the previous one” [1]. When the process is executed, it would lead for a definite number of output results. An example of algorithm used to identify disease cohort in this dissertation is presented in Figure 23, page 270. When this algorithm is executed on data of individual patient, the output is either ‘patient is included in the cohort’ or ‘patient is not included in the cohort’. ANKYLOSING SPONDYLITIS: “A chronic inflammatory disease of unknown origin, first affecting the spine and adjacent structures and commonly progressing to eventual fusion (ankylosis) of the involved joints … The disease primarily affects males under 30 years of age” [2].  ANTAGONIST: “Any agent, especially a drug or hormone, that reduces the action of another agent, the agonist” [3]. Antagonists work though several possible mechanisms: (a) combination with the substance being antagonized, (b) production of an opposite effect through a different receptor, (c) competition for the binding site of an intermediate that links receptor activation to the effect observed or (d) interference with other events that follow receptor activation [3].  BERNOULLI VARIABLE: “A variable having only two possible values (e.g. on or off, 0 or 1)” [1]. BIAS: “Systematic deviation of results of inferences from truth, or processes leading to such deviation” [1].    xvi BIOLOGIC DISEASE-MODIFYING ANTIRHEUMATIC DRUGS (BIOLOGICS): New antirheumatic drugs that target the dysfunctional immune system by using the body's own chemical mediators (manufactured by recombinant DNA technology) to reduce the immune mediated inflammatory response [4].  CENSORING (RIGHT): See RIGHT CENSORING. CLINICAL SIGNIFICANCE: “Importance, relevance or meaning for the care of individuals, who may be – in clinical research – patients. A difference in the size effect that is considered to be important (e.g. , by a professional) in medical decisions, regardless of the degree of statistical significance” [1]. CLUSTERING (in analysis): A method of analysis of data, in which the observations are divided into groups using several possible scemes (e.g. hirarchial). The observation in each cluster (group) are more similar to each other than observations in different clusters.  COMBINATION THERAPY: “The simultaneous administration of more than one drug to treat a single disease” [5]. COMPARATIVE EFFECTIVENESS: “A type of health care research that compares the results of one approach for managing a disease to the results of other approaches. Comparative effectiveness usually compares two or more types of treatment, such as different drugs, for the same disease” [6].   COMPARATIVE PERSISTENCE: A relative measure used to compare the persistence  (duration of treatment) between two groups of patients, usually treated with two different drugs from the same therapeutic group.  COMPLIANCE: “The extent to which a patient acts in accordance with the prescribed interval and dose of a dosing regimen” [7]. Compliance reflects the intensity of executing the prescribing recommendations [1]. CONFOUNDER: A factor that may explain or produce all or part of the association observed between the exposure and the outcome. The three conditions for a factor to become a confounder are: (a) it is a risk factor for the outcome (among the unexposed); (b) it is associated with the exposure of interest among the population from which the cases are derived and (c) it is not an intermediate step in the causal pathway between the exposure and the outcone [8]. CONFOUNDING BY INDICATION: “A type of confounding bias that occurs when a symptom or sign of a disease is judged as an indication (or contraindication) for a given therapy and therefore is associated with both the intake of a drug or a medical procedure (or its avoidance) and with a higher probability of the outcome… Cofounding by indication stems from an initial lack of similarity in the prognostic expectations of treated and untreated subjects” [1].    xvii CONFOUNDING BY SEVERITY: A type of confounding by indication bias that occurs when patients with the worst prognosis are assigned the most effective drug [8-12], and as a result, effectiveness is biased toward the null (no effect). CONFOUNDING: “The distortion of a measure of the effect of an exposure on an outcome due to the association of the exposure with other factors [confounders] that influence the occurrence of the outcome” [1].  CONTAMINATION, TREATMENT:  see TREATMENT CONTAMINATION  CROHN’S DISEASE: “A chronic inflammatory bowel disease of unknown origin, usually affecting the ileum, the colon, or another part of the gastrointestinal tract. Diseased segments may be separated by normal bowel segments, which give it the characteristic skip lesions” [2].  CYTOKINE: “Any of numerous small proteins released from a variety of cell types that affect cell behaviour. Cytokines can influence the cells releasing them or nearby cells; in some cases they can enter the bloodstream to influence distant cells. Cytokines are crucial to many aspects of cell proliferation, differentiation, migration, and function, and play a central role in immune responses and inflammation” [13].  DAYS-SUPPLY: “The number of days a prescription is intended to last. It is calculated by dividing the number of doses in the prescription by the number of doses per day” [14].  For the propose of this thesis ‘days-supply’ is defined as the number of days between the current dispensing and the expected next dispensing.   DIRECT MEDICAL COSTS (p. 217): “The cost of a defined intervention and all follow-up costs for other drug and health care interventions in ambulatory, inpatient, and nursing care”  [15]. DRUG PERSISTENCE: Duration of treatment with a specific drug. Patients who discontinued this drug or switched to another drug within the same therapeutic group are no longer persistent [16]. See also TREATMENT PERSISTENCE. DRUG: In this thesis the term drug refers to therapeutic drug which is any chemical substance used for medical treatment, cure or prevention of disease. Synonyms include medication and therapeutic regimen. DRUG-FREE INTERVAL: In analysis of administrative data, a drug-free interval is a gap in the treatment continuum during which the patient is supposedly not exposed to the drug. It is usually defined from the time the dispensed days-supply was exhausted until a new dispensing occurs. Also called grace period or permissible gap [17,18].  EFFECTIVENESS: “A measure of the extent to which a specific intervention, procedure, regimen or service, when deploy in the field of the usual circumstances, does what it is intent to do for a specific population” [1].     xviii EFFICACY: “ The extent to which a specific intervention, procedure, regimen or service,  produces a beneficial result under ideal conditions... Ideally, the determination of efficacy is based on the results of randomized controlled trial” [1].  FAIR PHARMACARE: A British Columbia drug reimbursement program, which started on May 2003. It covers the cost of most drugs for residents who are eligible based on net taxable family income criteria. See also PHARMACARE. HAZARD:  “A theoretical measure of the probability of occurrence of an event per unit time at risk; e.g., death or new disease, at a point in time, t, defined mathematically as the limit, as t approaches zero, of the probability that an individual well at time t will experience the event by t + t, divided by t” [1].  HETEROGENEITY: “(1) Used in a general sense to describe the variation in, or diversity of, participants, interventions, and measurement of outcomes across a set of studies, or the variation in internal validity of those studies. (2) Used specifically, as statistical heterogeneity, to describe the degree of variation in the effect estimates from a set of studies. Also used to indicate the presence of variability among studies beyond the amount expected due solely to the play of chance” [19]. IMMUNOMODULATOR: “Any agent that alters the extent of the immune response to an antigen, by altering the antigenicity of the antigen (i.e. the capacity of an agent to stimulate the formation of specific antibodies to itself) or by altering in a nonspecific manner the specific reactivity or the nonspecific effector mechanisms of the host” [3].  IMMUNOSUPPRESSANTS, IMMUNOSUPPRESSANT DRUGS: “Drugs that suppress the activity of the immune system”. These drugs are used in treatment of autoimmune diseases such, as rheumatoid arthritis [20].  INDEX DATE: A date that serves to guide, point out, or otherwise facilitate reference. In the context of the current research, the index date is defined as the date of the first dispensing of the of any TNFα antagonist drug.  INDIRECT MEDICAL COSTS (page 217): “All costs to the national economy of the society due to productivity loss. Indirect costs can be due to decreased efficiency, total absence from work through an illness, or premature death” [15]. INTERDOSE INTERVAL: The period (recommended or actual) between two administrations of the same drug. INTERMITTENT DOSING SCHEDULE: The drug is administered less frequently than once a day [21]. I-SQUARE QUANTITY: “A measure used to quantify heterogeneity. It describes the percentage of the variability in effect estimates that is due to heterogeneity rather than sampling error (chance). A value greater than 50% may be considered to represent substantial heterogeneity” [19].    xix JUVENILE RHEUMATOID ARTHRITIS (JUVENILE IDIOPATHIC ARTHRITIS): “A disease affecting the larger joints of children less than 16 years of age and often accompanied by systemic manifestations. As bone growth in children is dependent on the epiphyseal plates of the distal epiphyses, skeletal development may be impaired if these structures are damaged” [2].  META-REGRESSION: “An extension of meta-analysis in which the relationship between the treatment effect and known confounders is modeled using weighted regression. In this way insight can be gained into how the outcome is related to the design and population studied” [22]. MISCLASSIFICATION: A type of error in obtaining the needed information; errors in the measurement of subjects. “For discrete variables (variables with only a countable number of possible values, such as indicators for sex), measurement error is usually called classification error or misclassification. Classification error that depends on the values of other variables is referred to as DIFFERENTIAL MISCLASSIFICATION. Classification error that does not depend on the values of other variables is referred to as NONDIFFERENTIAL MISCLASSIFICATION” (page 138 [23]).  MONOCLONAL ANTIBODY: “An antibody produced by genetic engineering techniques from a cell clone (i.e. numerous identical cells originally derived from a single parent cell) and therefore consisting of a single type of immunoglobulin” [20].   MONOTHERAPY: Treatment of a disease by a single drug. NONSTEROIDAL ANTI-INFLAMMATORY DRUG (NSAIDs): “Any one of a large group of drugs used for pain relief, particularly in rheumatic disease associated with inflammation. NSAIDs act by inhibiting the enzymes controlling the formation of prostaglandins, which are important mediators of inflammation” [24]. OSTEOARTHRITIS (page 225): “A degenerative disease of joints resulting from wear of the articular cartilage, which may lead to secondary changes in the underlying bone. The joints are painful and stiff, with restricted movement. The condition may be primary or may result from abnormal load to the joint or damage to the cartilage from inflammation or trauma” [24].  PERMISSIBLE GAP: See DRUG-FREE INTERVAL. PERSISTENCE: See TREATMENT PERSISTENCE, DRUG PERSISTENCE. PHARMACARE (BRITISH COLUMBIA): The provincial Ministry of Health drug benefit program that pays for all or part of the cost of eligible prescription drugs and designated medical supplies through seven drug plans - the largest is the Fair PharmaCare plan, which is income-based [25]. See also FAIR PHARMACARE.    xx PHARMANET (BRITISH COLUMBIA): A province-wide database that links all pharmacies to a central set of data systems. Every prescription dispensed in the province is entered into the database. PharmaNet is administered by the Ministry of Health and the College of Pharmacists of British Columbia with the goals of supporting prescription claim processing and improve prescription safety [26]. PRIMARY INEFFICACY (p. 255): Lack of response, the drug never produced the desired effect. See also SECONDARY INEFFICACY. PSORIASIS: “A common chronic skin disorder characterized by circumscribed red patches covered by thick, dry silvery adherent scales that are the result of excessive development of epithelial cells. Exacerbations and remissions are typical” [2].  PSORIATIC ARTHRITIS: “A form of arthritis associated with psoriatic lesions of the skin and nails, particularly at the distal interphalangeal joints of the fingers and toes” [2]. REGISTRY: “The file of data concerning all cases of a particular disease or other health- relevant condition in a defined population such that the cases can be related to a population base” If cases are regularly followed up, information on remissions, exacerbations, incidence, prevalence and survival can be obtained [1].  RELIABILITY: “The degree of stability exhibited when a measurement is repeated under identical conditions. Reliability refers to the degree to which the results obtained by a measurement procedure can be replicated. Lack of reliability may arise from divergences between observers or instruments of measurement or instability of the attribute being measured” [1]. REMISSION: “A lessening in the severity of symptoms or their temporary disappearance during the course of an illness” [24]. Remission may be spontaneous or the result of therapy.  RIGHT CENSORING: “Loss or attribution of subjects from a follow up study, the occurrence of the event of interest among such subjects after a specific time when it is known that the event of interest had not occur; it is not known, however if or when the event of interest occurred subsequently” [1] . ROBUSTNESS: “A property of a statistical test or procedure that confers to it a certain degree of insensitiveness to departure from the assumptions from which it is derived” [1]. SECONDARY INEFFICACY (p. 255): Loss of efficacy in patients who previously responded to the treatment. See also PRIMARY INEFFICACY. SELECTION, DRUG SELECTION: In this dissertation, drug selection is defined as either (a) the physician decision as to which of the available therapeutic classes to use, following the decision to initiate pharmacological treatment for the disease; or (b) the decision as to which of the available drugs to prescribe, following the decision to prescribe a specific therapeutic class.    xxi SENSITIVITY ANALYSIS: “A method to determine the ROBUSTNESS of an assessment by examining the extent to which results are affected by changes in methods, models, value of unmeasured variables, or assumptions” [1]. SIGNIFICANCE, CLINICAL: See CLINICAL SIGNIFICANCE. SIGNIFICANCE, STATISTICAL: See STATISTICAL SIGNIFICANCE.  SPECIAL AUTHORITY PROCEDURE (pAGE 269): The British Columbia PHARMACARE procedure to grant coverage for drugs from several categories (limited coverage drugs, alternatives to Reference Drugs, alternatives to Low Cost drugs, some psychiatric drug and drugs included in the Alzheimer's Drug Therapy Initiative). Requests for Special Authority coverage must be submitted by the health care practitioner through a special application process. Authorization has to be granted before the plan provides coverage for the drug. The actual coverage is based on PharmaCare rules, including any deductible requirement [27]. STATISTICAL SIGNIFICANCE: “A statistical property of an observation or an estimate that is unlikely to have occurred by chance alone” [1]. SWITCHING (treatment): In the context of this dissertation, switching is the dispensing of a different drug within the same therapeutic class that occurred any time during the study period or a specified time period after a dispensing of the drug of interest. THERAPEUTIC CLASS: Therapeutic class is a term used to describe a collection of drugs that are used to treat the same disease and have similar pharmacological target. Multiple classification systems exists for drugs (e.g., the World Health Organization (WHO) system [28] and the American Society of Health-System Pharmacists’ Pharmacologic-Therapeutic Classification [29]).  TREATMENT CONTAMINATION (in epidemiology) (p. 25): “The situation that exists when a population being studied for one condition or factor also possesses other conditions or factors that modify results of the study” [1]. TREATMENT PERSISTENCE: The duration of time in which the patient is treated with the drug(s). It is measured from initiation to discontinuation of the treatment [7]. Treatment persistence can be categorized as DRUG PERSISTENCE, regimen persistence (any change in dosing or adding on drugs is considered discontinuation) and therapy persistence (switching is not considered discontinuation) (Table 1, page 24) [30]. ULCERATIVE COLITIS: “A chronic, episodic, inflammatory disease of the large intestine and rectum. It is characterized by profuse watery diarrhea containing varying amounts of blood, mucus, and pus” [2].    xxii ACKNOWLEDGEMENTS Achieving this important milestone has been made possible thanks to support from numerous individuals. First, I would like to thank my supervisor, Dr. Colin Dormuth, for his continuous support and guidance when needed. I am forever in his debt for giving me the confidence and freedom to explore my research interests. Colin,I am grateful for the time you have reserved for me in your busy schedule to discuss problems. Your constant advice has helped me keep focus and complete the research. The help, support and patience of my cosupervisor, Dr James Wright, has also proved crucial. His advice and unsurpassed knowledge of systematic review methodology helped me reach this point. I am grateful for his insights through the difficulties that arose during the research period. This thesis would not have been possible without the continuous encouragement of Dr. Ken Bassett. Not only did he inspire me to study the treatment presistance, Ken also helped me with interpretating the results and reviewing this dissertation. His contribution to my journey has been invaluable and enlightening. Thank you, Ken. I would also like to show graditude to my supervisory committee members, Drs. Alan Brookhart and Hugh Freeman. Thank you for the fruitful and enjoyable discussions, and for statistical and clinical insights into the subject studied.  Greg Carney and Richard Morrow, from the Pharmacoepidemiology Group (PEG), shared their time and SAS programming expertise with me. I deeply appreciate your help, which proved critical to solving many of the challenges encountered during the conduct of statistical    xxiii analyses. In addition, I owe my deepest gratitude to Vijaya Musini for introducing me to systematic review methology and sharing her knoweledge with me. I would also like to thank Ciprian Jauca, Christopher Adlparvar and Doug Salzwedel for their great collaboration and endless support, Tarita Miller for SAS program execution support, Stephen Adams for retrieving articles, and Carolyn J Green for editing. Finally, the patience of my family through my years of studies has been essential to the completition of this journey.To my father, Gidon, who was my first teacher, thank you for encouraging me to pursue lifelong learning. To my husband, Herbert, and my children, Keren, Ben, and Noa, who were a constant source of love, concern, and strength all these years, thank you.+De         xxiv DEDICATION    This thesis is dedicated to my late mother, Nurit, whose love and care has continued to motivate and inspire me. You taught me to never to quit and never to lose hope.    1 CHAPTER 1:  INTRODUCTION  This thesis explores comparative persistence of different tumour necrosis factor alpha (TNFα) antagonist drugs in patients with rheumatoid arthritis (RA). The thesis focuses on methodological aspects of measuring persistence in these patients, as well as the influence of various characteristics of physicians who provide them with medical care.  RHEUMATOID ARTHRITIS is a common chronic inflammatory disease that causes an approximately 1.5 fold relative increase in premature mortality and decreased life expectancy by 3 to 10 years compared with the general population [31], progressive morbidity and disability [32-36] and excessive economic burden [37,38]. Additional information on the disease is presented in Appendix A, Section A.1, page 215. The TNFα ANTAGONISTS - infliximab, adalimumab and etanercept - are a novel therapeutic class indicated in a variety of inflammatory diseases, including RA. The TNFα antagonists were the first immunomodulating drug agents targeting cytokines, which are themselves modulators of inflammation. The drugs block the physiological action of the cytokine TNFα, which performs an important role in the pathogenesis of inflammation and of RA. Although the three individual TNFα antagonists are considered to belong to a single therapeutic class, the class is diverse in pharmacological and pharmacokinetic characteristics [39-43]. These differences may affect therapeutic benefits and harms. Additional information on the drugs is presented in Appendix A, Section A.2, page 223. Treatment with TNFα antagonists does not cure patients [44], but rather treatment reduces symptoms, possibly induces remission, and prevents complications over the long term. TNFα    2 antagonists are considered by most physicians to benefit RA patients; however, the relative effectiveness of the three drugs is still unclear, mainly due to lack of head-to-head randomized controlled trials (RCTs) and a limited number of long-term observational studies. Additional information on TNFα antagonist treatment is presented in Appendix A, Section A.3, page 241. TREATMENT PERSISTENCE is defined as the duration of time that elapses between the date that pharmaceutical treatment is initiated to the date it is discontinued [7]. As former American Surgeon General C. Everett Koop is often quoted “Drugs don't work in patients who don't take them”, which means that therapeutic benefit depends on administration of the drug. Conversely, drug dispensing, as a measure of willingness to take a drug can be a measure of overall long-term harm-benefit balance of drug therapy, particularly in patients with alternative available therapies and chronic noncurable diseases [45]. Theoretically, patients will continue to adhere to a drug regimen as long as they experience or perceive a benefit to persisting and do not experience an unacceptable amount of harm. Drug persistence is defined as the duration a patients is exposed to a specific drug, as compared to COMPLIANCE, which defines the intensity of executing the prescribing recommendations (e.g., the proportion of doses taken as prescribed over a period of time).  Pharmacoepidemiology consists of four areas of research: (a) efficacy, the performance of the drug in ideal setting; (b) safety; (c) effectiveness, the performance of a drug in routine clinical setting and (d) drug utilization, which is “an eclectic discipline, integrating descriptive and analytical methods for the quantification, the understanding and the evaluation of the processes of prescribing, dispensing and consumption of medicines and for the testing of interventions to enhance the quality of these processes” [46]. Figure 1 presents the porpuse of estimating    3 treatment persistence within the framework of pharamcoepidemiological research. Treatment persistence research provides evidence of three of the areas of research: first, it is acceptable measure of drug utilization. In addition, it could generate signals on safety concerns, if persistence is diminished due to high risk of harm (adverse events). Lastly, since it is a measure that balances benefit and harm caused by the treatment, it can provides evidence on the overall effectiveness/risk of the drug . COMPARATIVE PERSISTENCE is a relative measure estimated by comparing the average (or median) duration of treatment between different drugs, usually from the same therapeutic class. As a relative measure, comparative persistence minimizes the effect of systematic errors such as misclassification of discontinuation.  That is, while different studies use different methods to ascertain drug discontinuation, individual studies usually use the same method for all drugs being compared.  Figure 1: Treatment persistence within the framework of pharmacoepidemiological research  Pharmacoepidemiology Efficacy Safety Effectiveness Drug utilization Treatment persistence    4 Comparative persistence is thought to mirror relative benefit-harm balance in chronic noncurable diseases, in that a drug with better persistence is considered more effective and/or safer. As such, comparative persistence estimates could be important to physicians and patients. Estimates of persistence and comparative persistence could also be important to policy makers and researchers. For policy makers persistence provides important information for cost-effectiveness estimates and for forecasting utilization. This could lead to restricted reimbursement policies, including sequence of treatment (e.g., a particular drug include as a benefit only as a second line). Comparative persistence is also important for conducting and interpreting effectiveness studies because differential persistence may bias the results. 1.1 Conceptual Framework and Previous Research Theoretically, four classes of factors could influence persistence and/or persistence estimating, and therefore are regarded as predictors of persistence (Figure 2). The first class is drug properties that could be divided into two main subcategories: (a) properties that determine the drug’s beneficial and harmful effects and (b) factors that influence persistence via treatment convenience and patient satisfaction. The drug pharmacological characteristics (structure, pharmacokinetics, mechanism of action, biological effects, immunogenicity, etc.) are a group of interlinking and relating properties that eventually determine the therapeutic effects and the benefit-harm balance of the drug treatment. A detailed description of these properties of the TNFα antagonists is presented in Appendix A, Section A.2.4, page 229. Other factor, such as    5 route of administration (Appendix A, Section A.3.6.1, page 262), cost and medications refill barriers could influence patient satisfaction and therefore could influence persistence. The second group of factors that could influence persistence is patient characteristics. This group includes (a) demographics, (b) disease characteristics/clinical presentation (disease severity, comorbidities, and previous treatment) and (c) psychological factors. Drugs may have different effect in different patient population. For example, greater benefit is often observed in the more severely ill and increased risk of harmful effects may be expected in in patients with certain comorbidities. Furthermore, patients may perceive disease severity and drug effects (benefit and harm) differently, hence may response differently with regard to treatment persistence in the same clinical situation [47]. A third class of factors that have the potential to influence persistence is physician characteristics. In the same clinical situation, different physician may response differently – some of them would recommend discontinuing treatment while other would recommend continuing the treatment. These individual patterns may be associated with prescriber demographics, education, adherence to clinical guidelines and experience, but also with psychological factors such as physician preferences or susceptibility to marketing policy. Further discussion on the effect of physician characteristics is presented in Chapter 5, page 128.    Finally, investigator or study design factors may influence persistence estimates. Intuitively, persistence, defined as the period of time from treatment initiation until discontinuation, is influenced by methods used for ascertainment of treatment initiation and discontinuation. A detailed discussion on different study design factor is presented in Chapter 2, page 12.     6 Treatment persistence Drug characteristics Benefit/harm profile Structure Pharmacokinetics Mechanism of action Biological effects Immunogenicity Factors that influence patients satisfaction Convinience Cost Patients characteristics Demographics Different effects of drugs Psychological factors Preceived disease severity Preceived benefit/harm Clinical presentation Different effects of drugs Physician characteristics Preference Education/ experience Demographics Adherence to clinical guidelines Speciality Study design factors Ascertainment of treatment initiation Ascertainment of treatment discontinuation Figure 2: Factors that influence treatment persistence (conceptual framework)    7 Similarily, comparative persistence, which could be regarded as the association between a drug (and its properties) and persistence could be influences by the remain third classes of factors: patient charateristics, physician characteristics and study design factors (Figure 3).  To date, drug persistence research has assumed that persistence is only determined by drug benefit-harm profile and patient characteristics. However, the consequences of different methods of measuring persistence, and the influence of different characteristics of the care- providing physicians, have not been studied except for one recently published study [48]. In that study, it was reported that significant clustering occurred around physicians of RA patients who discontinued TNFα antagonists due to lack of efficacy. This study is presented in detail in the Discussion Section of Chapter 5 (Section 5.5.1.3, page 153). Persistence and comparative persistence estimates for TNFα antagonists in RA vary in direction, significance and magnitude [49,50]. The reasons for these differences have not been previously studied. This thesis provides new knowledge of TNFα antagonist persistence and comparative persistence, as well as new and useful insights into the methodologies used to make these measurements and seek explanations for the observed differences in comparative persistence estimates.    8 Figure 3: Factors that influence comparative persistence (conceptual framework )    Comparative persistence Patients characteristics Physician characteristics Study design factors Different effect of different drugs Psychological factors Demographics Clinical presentation Preceieved disease severity Preceieved benefit/harm Psychological factors Education/ experience Ascertainment of initiation Ascertainment of discontinuation    9 1.2 Thesis Goals The main goal of this thesis is to compare persistence between three TNFα antagonists in RA patients: infliximab, adalimumab and etanercept. Additional goals are to 1) identify patient characteristics and physician prescribing habits that influence persistence, 2) identify confounding factors that bias estimates of comparative persistence, and 3) assess the impact of study design methodology on estimates of comparative persistence, specifically methodologies used to ascertain treatment discontinuation. In Chapter 3, we identify patient characteristics and additional factors that predict persistence and confound comparative persistence estimates. The influence of study design methodology was studied primarily in Chapter 4, in which we present the consequence of increasing the length of drug-free interval used in discontinuation ascertainment on estimates of comparative persistence. In Chapter 5 we study the effect of physician prescribing on persistence and comparative persistence.  Identifying predictors of persistence will help improve and standardize methodologies used in future research. It will also provide insights into pharmacoepidemiologic research using administrative health care data.    10 1.3 Overview of Thesis Chapters The thesis consists of six chapters organized in a manuscript-based format. Versions of Chapters 3, 4, 5 are written for publication as separate manuscripts.  Chapter 2 summarizes essential background on the methodologies used in estimating comparative persistence. It consists of two parts. The first part introduces general challenges in measuring persistence and comparative persistence with drugs. It also summarizes the various methodological approaches used in previous studies. The second part of the chapter presents the rationale behind the methods used in this thesis to study comparative persistence with TNFα antagonist in RA patients using administrative health care databases.  Chapter 3 is a population-based study that estimates comparative persistence with TNFα antagonists RA patients in British Columbia. The effects of various patient and physician characteristics are explored. Chapter 4 is a population-based study builds upon the work in Chapter 3, by critically examining the effect on comparative persistence estimates of using different algorithms to ascertain drug discontinuation. Specifically, it examines the effect of the length of the drug-free interval used in an algorithm, which is an exogenous factor chosen by the investigator, on the direction, significance and magnitude of hazard ratios for drug discontinuation. The study also describes patterns of reinitiating the same drug after a temporary interruption (30-180 days) in persistence.  Chapter 5 explores the extent to which persistence reflects prescriber rather than patient- characteristics. Specifically, Chapter 5 estimates the effect of prescriber preference for a    11 prescribed drug, where the preference is estimated based on previous TNFα antagonist prescribing habits. The study also describes patterns of switching to a second drug. Supplementally, it explores possible confounder effects of prescriber preference for the prescribed drug on estimates of comparative persistence.  Chapter 6 is a concluding chapter that summarizes and organizes the main findings and relevance of the thesis. It also recommends specific topics for future research.     12 CHAPTER 2:  METHODOLOGICAL ISSUES IN ESTIMATING PERSISTENCE This chapter summarizes essential background on the methods used to estimate comparative persistence in general and in this thesis specifically. The first part provides general methodological considerations in estimating persistence and comparative persistence. We summarize and discuss the use of comparative persistence methods in the context of research conducted by other researchers, and in relation to the actual outcomes measured. The second part presents the methodological framework used in this thesis to study comparative persistence with TNFα antagonists in RA patients. Data from the British Columbia administrative health care databases are used. We introduce some of the major challenges in the analysis, discuss possible solutions, and present a selected methodological approach along with a rationale for our approach.  2.1 General Methodological Background  The general methodological background consists of a literature review that provides an overview of terminology and analytic issues. First, we present the terminology used (in the literature and in this thesis) to describe the degree to which patients follow prescribed dosing recommendations. Next, we introduce some of the major challenges in measuring treatment persistence, specifically ascertainment of treatment initiation, discontinuation and the date of discontinuation. We present these challenges in light of the data source used: (a) clinical data    13 from clinical registries or medical charts and (b) administrative health care databases that include data on prescription drug dispensing claims. Lastly, we summarize the essential background on the methods used to estimate comparative persistence.  2.1.1 Terminology  The degree to which patients follow prescribed dosing recommendations is an important determining factor of therapeutic outcomes. Not taking drugs as prescribed could lead to poor outcomes because the drug is not available in the optimal quantity to affect the patient’s disease process [51-53]. Conversely, lack of therapeutic effect or the occurrence of side effects may interfere with adherence with dosing recommendations. Measures of the degree to which patients follow prescribed dosing recommendations reflect an association with final health outcomes rather than a straightforward causal relationship. The current medical literature is characterized by a lack of clarity on the use of the terms persistence, compliance and adherence, where often these words are used interchangeably [7]. PERSISTENCE reflects “the act of conforming to a recommendation of continuing treatment for the prescribed length of time” [7]. Persistence means that a patient continues to take a drug (or in an analysis of administrative data - refill the prescription) while DISCONTINUATION or nonpersistence means that the patient stopped the drug (or stopped refilling) at some point. Persistence can be documented as a continuous or a Bernoulli variable. Most commonly it is reported as a continuous variable and referred to as the duration of time from initiation to discontinuation of therapy. Basically persistence reflects the number of days, months or years    14 for which therapy was actually administered (Figure 4) [7] although some exceptions are applied. One exception occurs with the use of administrative data, since only dispensing claims are available to approximate actual drug administration. In an analysis of administrative data, therefore, persistence usually reflects the number of days the drug was available. In drugs with an intermittent dosing schedule (i.e., less than once daily) persistence is usually defined to reflect the number of days the patient was experiencing the therapeutic effect (e.g., until the first missed dose). Alternatively, persistence can be reported as a Bernoulli variable (i.e., whether or not a patient is persistent) at the end of prespecified period after the initiation of treatment [54]. For example, a patient is considered persistent at one year if the drug has not been discontinued by that time. Synonyms for treatment persistence used in  Figure 4: Persistence, adherence and compliance          time Drug initiation Drug discontinuation Adherence =        +         Persistence = days on drug  Compliance = % taken as prescribed    15 the scientific literature include ‘duration of treatment’, ‘drug survival’  (as imprinted by Wolfe [45]), ‘treatment continuum’, ‘treatment continuation’ (all for continuous presentation) and  ‘drug retention rate’1 (for binomial presentation).  COMPLIANCE is defined as the extent to which a patient acts in accordance with a prescribed interval and dose [7]. It captures the intensity of the treatment compared to the recommended dosing while persisting. Noncompliance means that the patient is late to take a dose as scheduled or recommended and misses as least one day of therapy. Usually compliance is measured over a period of time and reported as a percentage of doses taken according to prescription (Figure 4).  Finally, in the literature the term ADHERENCE is often used interchangeably with the term compliance to describe the extent to which a patient acts in accordance with the prescribed interval and dose. However, the World Health Organization (WHO) emphasized that adherence also requires the patient’s agreement to the recommendations [55]. In this thesis, ADHERENCE is used as a comprehensive term that includes three types of errors in dosing: nonacceptance, poor execution (compliance) and early discontinuation (persistence) (pp. 368-9 [56]). Our use of adherence encompasses both intensity (following the recommended interval and dose, compliance) and duration of time (persistence).                                                  1  Drug retention time, on the other hand, is used to describe a pharmacokinetics property of drugs, which is closely related to bioavailability and is affected by the half-life of the drug. This term may be confusing, hence will not be used in this thesis.    16 2.1.2 Measuring Drug Persistence Two types of studies characterize the literature on treatment persistence. The first type consists of original persistence studies conducted using various methodologies. The methods used varied according to the form of the persistence variable that was used (continuous or Bernoulli) and on the type of data source that was analyzed (clinical or administrative). The second type of academic publication does not report original persistence research but rather reviews methods used in persistence studies or suggests methodological guidelines on how to best conduct these studies [57-59]. In this section (Section 0) we discuss the methods most commonly used to estimate persistence as a continuous (Section 2.1.2.1) or a Bernoulli (Section 2.1.2.2) variable. We complete the discussion by presenting methodological and contextual issues in ascertaining the treatment initiation and discontinuation parameters that are required to estimate persistence. 2.1.2.1 Persistence as a Continuous Variable Treatment persistence is commonly regarded as a continuous outcome variable that can be estimated using time-to-event data. Persistence is measured from the date on which the patient was first exposed to the drug (often defined as the index date, Section 2.1.2.3, page 19) until the date on which treatment is discontinued (Section 0, page 24). Patients are followed from the index date until they discontinue the drug or until the end of a prespecified follow-up period. Some patients are still on the treatment at the end of the prespecified follow-up period; hence, their data are considered to be (right) censored. Survival analysis methodology is used    17 to account for right censoring. This method allows maximum use of available data by accounting for patients who were censored and utilizing their known duration of treatment.  The main method used in studies of TNFα antagonists is Kaplan Meier product limit approach that allows direct estimation of persistence in one or more patient groups. Persistence is reported as a median value2  with either a confidence interval or an interquartile range (IQR) provided to confer the degree of sampling variability. Kaplan Meier is considered to be a nonparametric approach, i.e., no assumptions on the distribution of persistence in the study population are made. Drug discontinuation3 is considered ‘failure’ [57] and persistence is ‘survival’. The Kaplan Meier approach is based on several assumptions [60]. First, study patients are assumed to be drawn from a random sample of the target population. The second assumption is that the observations are independent of each other, meaning that choosing any one patient in the population does not affect the chance of choosing any other patient. Third, the index date (treatment initiation, cohort entry) and the treatment discontinuation date (failure) are both assumed to be defined and ascertained consistently during the study period. Fourth, average persistence is assumed to be constant during the course of the study. Finally, the time of                                                  2  Uncommonly, treatment persistence is reported as mean, which is a biased estimate. Mean persistence is estimated as the area under the survival curve, and based upon the entire range of data. The mean is underestimated when the longest observation is censored [476,477]. In addition, since time-to-event data is right-skewed, median is a better measure of central tendency [478].  3  Also called ‘termination’, ‘withdrawal’, ‘ending’ or ‘stopping’ the treatment    18 censoring is assumed to be unrelated to persistence, meaning that, on average, the persistence of censored patients is identical to the persistence of noncensored patients. This last assumption is especially critical to the validity of results when analyzing data in which many patients are censored. This assumption can be assessed using information on patients who were lost to follow-up. If the reason for leaving the study relates to persistence, then the estimated persistence is biased.  2.1.2.2 Persistence as a Bernoulli Variable Treatment persistence can also be analyzed using a Bernoulli outcome variable, depending whether or not the patient was persistent at specific time points date (e.g., anniversaries). For example, when estimating one-year persistence, patients who are considered persistent are assigned the value 1 to the outcome variable, and patients who are not persistent are assigned a value of 0.   There are multiple approaches used to ascertain persistence at specific time points [59]. The most common approach is to consider patients as persistent until treatment discontinuation. Using the same example, a patient is considered persistent (outcome variable=1) if discontinuation was not observed during the first year of follow up, and nonpersistent (=0) if discontinuation was observed during this year. Similar to the continuous variable approach (Section 2.1.2.1, page 16) classifying the Bernoulli variable as 1 or 0 requires observation of the index date and discontinuation date as described below (Sections 2.1.2.3 - 0, from page 19).  Alternatively, a Bernoulli persistence variable can be assigned as 1 or 0 using measures of compliance (intensity of following the dosing recommendation) at a specific time interval [59].    19 A patient is persistent as long as the compliance measured is higher than a predefined threshold. For example, a threshold of three refills in the first year of treatment can be predefined as providing sufficient evidence of compliance. A patient who refills 0-2 prescriptions during the year is therefore considered nonpersistent and the outcome variable is assigned a value of 0. A patient who refills three or more prescription during the year is considered persistent and the outcome variable is assigned a value of 1. 2.1.2.3 Ascertainment of the Index Date The index date is defined as the date of treatment initiation. Depending on the data source (clinical data or administrative data) and route of administration, the index date is defined as the date of the first prescription, dispensing or administration of the drug. In previous observational studies of TNFα antagonists in RA patients using administrative data [61-64], the index date was defined as the date of the first dispensing or the first insurance claim (usually but not always these dates are identical) in the study period. This definition assumes that the drug is administered immediately after the dispensing. In analyses of clinical data, the index date is often defined to be the date of first prescription. Examples include studies of TNFα antagonists in RA patients [65,66] and studies of other diseases using the United Kingdom General Practice Research Database [67]). Alternatively, when a drug is administered intravenously (e.g., infliximab) the index date can be defined as the first administration date [65,68,69].  A valid estimate of treatment persistence require new-user study design. In this design, patients A new-user design is restricted to patients with a minimum period of nonuse (washout or run- in) prior to treatment initiation [70], a period in which patients are not exposed to the drug    20 studied. In the absence of run-in period, the measured duration of treatment could be underestimated, since the patients might have been exposued to the drug before the initiation date.  Bias may be introduced as a consequence of the methods used to ascertain the index date. First, persistence may be overestimated in analysis of administrative data as the result of using dispensing data when there is a delay or failure in administering the drug. In practice, administration is assumed to occur immediately follow the prescription or dispensing and therefore this source of overestimation is viewed as minimal and insignificant. Second, differential misclassification may occur across study groups. This may occur, for example, when estimating persistence of several drugs from clinical data where different methods are used to ascertain the index date. Specifically, differential misclassification may have occurred in comparing infliximab to adalimumab or etanercept when the index date was ascertained from the first intravenous administration for infliximab and the date of prescribing for adalimumab or etanercept (self-administered) [65,71]. In this example, persistence with the latter drugs may have been overestimated compared to infliximab.    21 2.1.2.4 Ascertainment of Treatment Discontinuation  Ascertainment of treatment discontinuation from medical records, prospective clinical databases and patient surveys is relatively straightforward for TNFα drugs. Study researchers usually ascertain discontinuation based on records kept by the care-providing physician or from patient self-reports. However the exact method is usually poorly reported or not provided [49]. Ascertainment of treatment discontinuation presents a major challenge in analyses of administrative databases of dispensing claims. In these analyses, ascertainment of discontinuation is usually based on gaps in exposure (or availability) of the drug [17]. Typically, discontinuation is ascertained based on a DRUG-FREE INTERVAL (also termed permissible gap or grace period) of 7-180 days [17], often 30-60 days [18] after the dispensed quantity for each refill was exhausted. An algorithm is generally based on summing the dispensed days-supply of each refill and the predefined drug-free interval. If no additional dispensing claim is recorded during this period, the patient is considered to be a discontinuer, and any further dispensing events of the drug are censored (Figure 5). In studies of persistence with TNFα antagonists in RA patients using administrative data, drug-free intervals of 30 [61- 63], 60 days [72] or 90 days [64] have been used.  The rationale of using a drug-free interval is to minimize underestimation of persistence. Several sources of bias are possible if discontinuation is assumed to occur when no refill occurs at the same date that the dispensed days-supply was exhausted. First of all, this zero-gap is very sensitive to compliance in that patients who experience even a temporary lack of compliance are defined as discontinuers in the analysis. Additional sources of bias include     22 Figure 5: Establishing discontinuation based on days-supply and drug-free interval  The figure presents two patients, who did not renew the drug once the dispensed days-supply of the third refill was exhausted. In patient A, the fourth dispensing event occurred after the predefined drug-free interval elapsed. Therefore, this patient is considered to be a discontinuer. In patient B, an additional dispensing event was recorded before the drug-free interval elapsed; hence, the patient is NOT considered to be a discontinuer.  errors in recorded quantity or days-supply dispensed, use of supplies that were previously dispensed and stockpiled, dose reduction and the use of samples or hospitalization [73,74]. Selection of a drug-free interval should be based on the pharmacology of the drug and the outcome of interest. However, the rationale for the length of a drug-free interval, if used, has only rarely been provided in published studies. The length of drug-free interval should be determined based on the pharmacologic properties of the drug and in the context of the treatment situation [7,59]. In studies of TNFα antagonists the duration of therapeutic effect has time  Days  supply dispensed time  Predefined  drug-free interval A. B. Days  supply dispensed Days  supply dispensed Days  supply dispensed Days  supply dispensed Days  supply dispensed Predefined  drug-free interval    23 been shown to significantly exceed the typical interdose interval in patients with RA. For example, the estimated median duration of effect on infliximab has been shown to be 16.5 weeks which is twice as long as the typical interdose interval of eight weeks [75]. The effect of adalimumab can be sustained in patients for four weeks to three months, which is many times longer than the typical interdose interval of two weeks [76]. This aspect of TNFα antagonist treatment is critically important because it suggests that using short drug-free intervals (e.g. 30 days) leads to underestimation of the duration of therapeutic response and patients may therefore be categorized as discontinuers even though they continue to benefit from the drug. Second, the length of the drug-free interval used in a study substantially depends on whether the research is measuring persistence or adherence. In studies of persistence, discontinuation is the event of interest. The chance that discontinuers reinitiate treatment with the same drug is minimized by selecting a longer drug-free interval. In studies of adherence, the researcher may be interested in capturing also temporary interruptions in persistence, after which the patient reinitiates the treatment. In these studies the outcome integrates both duration and intensity of treatment, and the event of interest is nonpersistence [77] or nonadherence. It is more readily ascertained by using shorter drug-free intervals.  The selection of the length of the drug-free interval is often challenging and there is no consensus on the optimal way to choose an interval any given context. It has been previously suggested that examining the sensitivity of persistence estimates to a varying drug-free interval is appropriate [59,73] or as suggested by Caetano, it is appropriate “to examine a spectrum of persistence according to different levels of intensity” [59].    24 Additional methods to ascertaining discontinuation are found in the literature. For example, treatment discontinuation was also ascertained based on multiples of the days-supply dispensed [17] or a specified time period (in days) after each dispensing event (Refill-Sequence Model) [17,59]. The last method was used in a recent study of TNFα antagonists in RA for example [78].  Three additional changes in the prescribed regimen are of interest in studies of persistence: addition of another drug, dose adjustment and switching, defined as the dispensing of a different drug within the same therapeutic class that occurred any time during the study period or a specified time period after a dispensing [17]. Halpern et al 2006 [30] suggested that measures of treatment persistence could be categorized to (a) drug persistence, (b) regimen persistence and (c) therapy persistence (Table 1). The most common approach is DRUG  Table 1: Three approaches to treatment persistence (adapted from Halpern 2006 [30]) Approach Definition Impact on persistence Discontinuation of the drug Addition of another drug Dose adjustment Switching to another drug Drug persistence Time on a specific drug Yes No No Yes Regimen persistence Time on a specific set of drugs without any change Yes Yes No Yes Therapy persistence Time on any drug for the disease Depends whether the patients was switched or not. No impact if the patient was switched to another drug No No No     25 PERSISTENCE (medication persistence), specifically patients who added-on drugs to initial therapy are usually considered to be persistent [16] and patients who switch away from their initial drug are considered by researchers nonpersistent (discontinuers) [16,61,62,64,79]. This is the approach used in this thesis, and therefore the term ‘drug persistence’ in used in the following text. Of importance is that switching or addition of new drug does not always indicate drug/treatment failure. If a new comorbidity has developed, for example, the patient may switch to or add a second drug that is indicated to both the disease of interest and the new comorbidity [30]. The Refill-Sequence Model that previously mentioned permits switching within the same therapeutic class or among drugs with shared indications and these patients are not considered to be discontinuers [59] and therefore is considered a therapy persistence approach. The approach to switchers depends on the study goal. When estimating comparative persistence, switchers should be considered discontinuers, to avoid cross-contamination by alternate drug exposures. However, if the interest is in comparing treatment approaches or clinical situations, then according to the ‘intention to treat’ approach switchers could be considered as to be persistent (i.e., therapy persistence approach, Table 1). Generally, dose adjustments (increase or decrease) do not reflect lack of compliance or persistence and therefore are not considered nonpersistence [30]. All the above methods require reliable and accurate recording of days-supply, which may be challenging especially in analysis of administrative data[59]. This issue is discussed in detail below (Section 2.2.4.1, page 47).     26 2.1.2.1 Ascertainment of the Date of Treatment Discontinuation  Once discontinuation is confirmed, how the date of discontinuation is ascertained varies depending on the data source. The date of discontinuation is the last date on which the patient has been exposed to the drug. Alternatively, in drugs with prolonged therapeutic effect such as  on which the patient has been exposed to the drug effect.  When clinical data is used to estimate persistence with TNFα antagonists, the date of discontinuation can be defined as the date of the last administered dose, the date of the decision to discontinue or the first missed dose. Unfortunately, the method used is often poorly described in published studies.  In analysis of administrative data, ascertainment of the date of discontinuation is more challenging. In administrative databases with dispensing claims, there are generally fields for recording the dispensing date and the dispensed days-supply. When considering a drug which is administered at least once a day, the end of days-supply as recorded is often used as the date of drug discontinuation. Nevertheless, this may overestimate persistence, since patients may not use the entire quantity dispensed. Treatment with TNFα antagonists presents a more complicated situation due to the varied and prolonged interdose interval4 (intermittent dosing schedule). The recorded days-supply may reflect the actual number of days until the next refill is expected or the number of days in which the drug is actually administrated (Section 2.2.4.1,                                                  4  Recommended dosing schedules for the TNFα antagonists in patients with RA are presented in Appendix A, Table 20, page 3.    27 page 47). Several approaches that can be used to ascertain the date of discontinuation and their limitations are presented below. 1. The date of the last refill – Using this approach the persistence could be underestimated and the effect on persistence with the TNFα antagonists may vary by the interdose interval.  2. The date on which the days-supply was exhausted, also known as the date of the first missed refill – A major challenges is to validate the recorded days-supply or establish the actual days-supply. This method is based on the assumptions that all of the dispensed quantity was administered and that no effect of the drug continues once the days-supply runs out. The last assumption may not hold in treatment with TNFα antagonists in RA patients, as discussed previously (Section 2.1.2.4, page 21).  3. The date of the end of the therapeutic effect - This method is not used in practice because estimating the duration of the effect is demanding and there is large variability between patients. When using a drug-free interval to ascertain discontinuation, the use of a date earlier than the end of the interval, such as the date on which the days-supply was exhausted, possess an additional complication. In this case, the investigator looks ahead in time and uses “future” data to ascertain an event that occurred in the past.  This approach is unacceptable in research of therapeutic outcomes, mainly since in this case it is impossible to suggest a causal relation, For example, an uncertainty would be arisen whether treatment discontinuation caused the outcome, or whether the “outcome” caused treatment discontinuation. When estimating treatment persistence, setting the discontinuation date to the end of days supply would robustly    28 results in estimates shorter by the length of drug free interval compared to a situation in which the discontinuation date is set to the end of drug-free interval. When using fixed length of drug free interval in studies of comparative persistence, the comparison should not be significantly biased, because the difference in persistence using the two methods is fixed for both drugs. 2.1.3 Comparative Analysis of Persistence Similar to persistence, estimating comparative persistence can be done using the Kaplan-Meier approach with a log-rank chi-square to test for differences between the persistence of any number of drugs. Alternatively, regression analysis can be used to investigate the comparative persistence of two or more individual drugs. In general, comparative persistence can be estimated using an additive risk model, and be expressed as differences in risk for discontinuation or using a multiplicative risk model, expressed as a risk ratio. Risk differences are typically more informative than ratios; nevertheless, most published analyses of comparative persistence were conducted using the multiplicative scale and Cox proportional hazard regression.  2.1.3.1 Cox Regression   In Cox proportional hazard regression, results are reported as hazard ratios for drug discontinuation. Discontinuation has an inverse relationship to persistence: better persistence on a drug can also be expressed as a lower risk of discontinuation. However, the size of the ratio is not proportional: a hazard ratio of two means that the drug of interest is associated with    29 a two-fold increased hazard of drug discontinuation, but the persistence with this drug is not necessarily half the persistence of the comparator drug. The two-fold increased hazard is interpreted as following: patients treated with the drug of interest and who were still persistent at time t (i.e., did not discontinue) have twice the chance of discontinuing at time t compared to patients treated with the comparator drug.  The multiplicative approach is especially attractive because researchers can adjust and study the effect of covariates. It also allows the baseline hazard to be unspecified. In addition, interpretation of the estimated coefficients is simple. The main limitation of this approach is the difficulty in interpreting the hazard ratios and associating them to persistence. Hazard ratios should be regarded as a measure that allows hypothesis testing, but ideally it should be used in combination with a measure of duration (such as median persistence) to adequately describe the size of the effect.  2.1.3.2 Aalen’s Additive Risk Model Alternatively, the additive risk model suggested by Aalen [80,81] is useful when risk difference is the main interest as opposed to relative risk. The hazard at any time is a sum of an unknown baseline hazard and a linear combination of covariate values. Estimation of the risk coefficients is based on a least-squares method. The additive risk model allows covariate    30 effects to vary with time. The disadvantages of this model are its infrequent use and its inappropriateness for use with some datasets5, in contrast to the Cox model [82].  2.2 A Methodological Frame for Estimating Comparative Persistence Using British Columbian Ministry of Health Databases This section presents the methodological framework we used to estimate comparative persistence of selected TNFα antagonists using British Columbia Ministry of Health databases. We introduce the data source and major methodological challenges that needed to be overcome to conduct the research for this thesis. To use administrative data intended for financial and management purpose to answer clinical questions presents major methodological challenges. One strategy is to create algorithms - computerized step-by-step lists of simple instructions. Algorithms were used to identify the study cohort, to calculate persistence of individual patients and their censor status6 and to assign values for the exposure and covariates.                                                   5  The dataset requirements are described in details in Cao’s dissertation [82] and are beyond the scope of this thesis. 6  The concept of censoring and its significance are discussed in Section 2.1.2.1, page 3.    31 2.2.1 Data Source: The British Columbian Ministry of Health Databases   The British Columbia Ministry of Health databases are useful powerful resource for conducting population-based health research, particularly because patient-level information from multiple databases is available . For this thesis, data records were accessed from the original British Columbia Health Data Warehouse platform for this research [83]. The new platform, HealthIdeas data warehouse services was designed to enhance security and privacy, using tiered zone based architecture. It has been in operation from 2010 and was not used in this research.  For the purpose of our study, we used a dimensionalized database platform7 that allowed a tailored access to various attributes of the following data, using a study-specific anonymized identification number: 1. Registration data (from January 1, 1986) - The dataset contains basic demographic data, such as age and sex, as well as data on registration in the provincial Medical Service Plan (MSP). First Nations and federal employee data were not accessible in the view of the dataset we obtained.                                                  7  A dimensionalized database is a specialized data storage facility that stores summarized data for fast and easy access. The summary is available across different dimensions of data such as location, time and individual patient. Users can quickly view large amounts of data as a value at any cross-section of the dimensions of interest.    32 2. MSP payment information data (from April 1, 1985) - Data included all medically required services provided by fee-for-service practitioners (mainly physicians, but also some additional health providers such as nurse practitioners, chiropractors, massage therapist, physiotherapists) in British Columbia, for British Columbia residents covered by the provincial MSP. Data was also available for British Columbia residents who obtained services in Quebec, the United States (U.S.) and other countries that were paid by MSP as well as services provided by fee-for-service practitioners billed at the regular rates, but paid for by third parties such as the Insurance Corporation of British Columbia (ICBC)8 and WorkSafe BC9. The dataset did not include services provided through alternative payment plans (salaried, sessional, and service agreement contracts), services for British Columbia residents who obtained medical services in other provinces and territories (except Quebec) and payments to physicians who perform WorkSafe BC expedited surgeries in public hospitals. Data on abortion (preganancy termination) procedures were also not included. The majority of billing records are submitted electronically by practitioners' offices to MSP. Diagnostic codes collected in the International Statistical Classification of Diseases and Related Health Problems version 9 (ICD-9) format (one to three codes per record) and considered accurate only to the 3rd digit level. MSP conducts routine audits/quality checks                                                  8  ICBC is a provincial corporation that covers expenses, including medical expenses, cause by bodily injuries in motor accidents to BC motorists. 9 WorkSafe BC (the Workers' Compensation Board of British Columbia) promotes workplace health and safety. In the event of work-related injuries or diseases, the board is responsible to health care benefits, compensation, return-to-work rehabilitation, and a range of other services [479].    33 on select fields. Diagnostic codes have been shown to be valid at the population level [84].  3. Hospital separation data (Discharge Abstract Databases, DAD) (from April 1, 1985) - Our access included information on the discharge, transfer and death of inpatients (and day surgery patients) from British Columbia acute care hospitals (British Columbia residents and non- British Columbia residents) as well as data on British Columbia residents who are admitted to a hospital in another province. Data on visits to emergency rooms/departments as well as outpatient services were not recorded in this file. Data on abortion procedures were not included as well. The format of the data changed significantly in 2001/02, as coding diagnoses switched from ICD-9 codes to ICD-10-CA10 codes, and procedures/interventions coding from CCP11 to CCI12. The centralized data processing of DAD records by the Canadian Institute of Health Information (CIHI) results in increased efficiency and standardization among the participating provinces. The CIHI performs an annual data quality study [85].  4. PharmaNet Data (from September 1995) - We obtained access to data on drugs dispensed in British Columbia pharmacies (prescription drugs only, excluding the prescriptions of hospital inpatients) and PharmaCare claims data.                                                  10  The ICD-10-CA is an enhanced version of ICD-10 developed by the Canadian Institute of Health Information for morbidity classification in Canada. 11  The Canadian Classification of Diagnostic, Therapeutic and Surgical Procedures 12  Canadian Classification of Health Interventions    34 Administrative health data, also known as claim data, are collected primarily for reimbursement, but usually contain some clinical data, such as diagnosis and procedures reimbursed [86]. Administrative data offer several advantages in research of therapeutic outcomes, especially in chronic disease. Data are collected in a systematic manner and usually are population based. Hence studies are quick, cheap and generalizable [87,88]. Administrative data provide estimates of effectiveness from a variety of clinical setting rather than efficacy and allow comparison of multiple alternative interventions beyond a placebo comparator as well as comparison of clinical strategies [86]. Large databases also allow research on rare and often delayed drug-related outcomes (benefit or harm) [86-88]. This type of data provides information on how the drug is dosed and applied in routine clinical practice [86]. In addition, administrative data is usually up-to-date and follow-up data is available before and after an event of interest. Finally, research using administrative health data prevents three types of bias that is present in other studies.  First, since informed patient consent is not required studies are less prone to selection bias [89-91]. Second, information recorded in the pharmacies is usually more accurate than self-reported drug use that may introduce reporting bias (especially for elderly, ill or deceased) [89,92]. Finally, since the analyses are nonintrusive, the research and the researchers do not influence the outcomes [92,93]. Participants and providers in RCTs may change their behaviour (Hawthorne effect13), and this change may affect the magnitude of the                                                  13  Hawthorne effect is “the effect (usually positive and beneficial) of being under study upon the person being studied; their knowledge of the study often influences their behavior” [1].    35 outcome for each group, and also may result in an inflated estimate of the effect size of the active comparator [94,95].  2.2.2 Creating the Study Cohort The British Columbia physician services database is primarily used for payment purposes, and therefore the diagnostic coding in that database is often unreliable. As a consequence, the first challenge in this research was to identify patients with RA that would comprise the study cohort. Ideally, an algorithm with high specificity is used to identify a study cohort (i.e., all patients included in the cohort have RA). Specificity of potential algorithms can be conducted in validating studies in comparison to clinical data. Since we had no access to clinical data, we were dependent on previous validation studies. The first step in developing an algorithm for our research was to design a disease algorithm in which patients diagnosed with RA could be identified. Prior to conducting the research and selecting the disease algorithm, we identify four different validation studies of different disease algorithms used on a variety of databases (Table 2). Two algorithms were identified that were based on a single recorded diagnosis of RA resulting in a sensitivity of 0.65-0.90 and positive predictive value (PPV) of 0.86-0.95.     36 Table 2: Published RA disease algorithms  Study Setting Algorithm examined Compared to Results Katz et al 1997 [96] 8 rheumatologic clinics in 3 U.S. states ICD-9 code of RA (714) used in Medicare Part B physician claim files Criteria by the American College of Rheumatology  Sensitivity 0.90 (95% CI 0.85-0.95) (N=342) PPV 0.95 (95% CI 0.92-0.98) (N=307) Losina et al 2003 [97] THR recipients N=922 ICD-9 codes of Medicare physician claims Hospital admission as recorded by physician Sensitivity 0.65 (95% CI 0.49-0.80).  PPV 0.86 (95% CI 0.73-0.99). Singh et al 2004 [98] Veterans hospital N=184 ICD code 714 for RA diagnosis in an administration hospital database Chart documentation of RA diagnosis by a rheumatologist on at least 2 visits at least 6 weeks apart Sensitivity 1.00 (95% CI n/a) PPV 0.66% (95% CI 0.59-0.73).  ICD code 714 for RA diagnosis in an administration hospital database plus synthetic antirheumatic drug 14  prescription fill  Sensitivity 0.85 (95% CI 0.80-0.90) PPV 0.81 (95% CI 0.75-0.87) Lix et al 2006 [99] Manitoba   16 different algorithms were examined Canadian Community Health Survey, cycle 1.1, collected between September 2000 and November 2001 [100] Table 7 of the report presents the accuracy of different algorithm examined varied between 0.05-0.11 in different algorithms.  PPV ranged                                                  14 The synthetic antirheumatic drugs are traditional short synthetic drugs from a variety of therapeutic classes used in treatment of RA. The term disease modifying antirheumatic drugs (DMARDs) or nonbiologic DMARDs have often been used in the literature to identify drugs from this heterogeneous group, regardless of lack of evidence that these drugs actually change the natural history of the disease. They include azathioprine, cyclosporine, hydroxychloroquine, leflunomide, methotrexate (MTX), penicillamine, sodium aurothiomalate, sulfasalazine and others.    37 Study Setting Algorithm examined Compared to Results between 0.63-0.81 For example, algorithm requires 2 RA diagnosis in 1 year have sensitivity of 0.05 and PPV of 0.77 CI – confidence interval;  ICD - International Classification of Diseases; n/a – not available; PPV – positive predictive value; RA – rheumatoid arthritis   38 For the purpose of evaluating the study hypothesis, we were interested in maximizing the PPV and minimizing the number patients with false RA diagnosis that were included in the cohort. Therefore we used similar, though not identical, a disease algorithm to that used in previous studies of RA patients in British Columbia [101-103]. RA patients were identified based on two outpatient visits with a diagnosis code of RA at least 60 days apart, or one hospitalization with a recorded discharge diagnosis of RA in three years prior to the index date. The list of ICD codes we used is presented in Appendix B, Table 28, page 271. We required two outpatient encounters with diagnosis of RA to increase the PPV and spaced at least 60 days apart to exclude transient joint disease or a diagnosis recorded as part of a diagnostic process in patients with other diseases. We regarded hospital diagnosis to be more accurate than outpatient diagnosis, since they are based on the discharge abstract, which is required for medical, and not payment, purposes.  In a second step, we designed an algorithm to exclude patients with diseases other than RA to increase the specificity of the study cohort. Treatment with TNFα antagonists is also indicated for Crohn’s disease, ulcerative colitis, juvenile RA, ankylosing spondylitis, psoriasis and psoriatic arthritis. Previous persistence studies of RA patients using administrative data have similarly excluded patients with other diseases indicated for TNFα antagonists including juvenile RA (patients younger than 18 years at index) [61,64,72,104] and Crohn’s disease [61,72].Our algorithm also excluded patients with either juvenile RA or Crohn’s disease. In a preliminary analysis not presented in this dissertation, we found that crude persistence (based on Kaplan Meier curves and log-rank test) was significantly different between RA patients and Crohn’s disease patients. Patients with both diseases presented persistence similar to patients with Crohn’s disease. Therefore, we excluded patients with at least one recorded diagnosis of    39 Crohn’s disease. We also routinely excluded young patients (under 18 years at index date), because they most probably had juvenile RA (also known as juvenile idiopathic arthritis) which is a similar, but not identical disease with different prognosis. In addition, patients with no data on the date of birth were excluded. Patients with recorded diagnosis of other indications approved for TNFα antagonist treatment were not excluded, due to small number of patients. Ankylosing spondylitis was approved in middle of 2005 and psoriasis, psoriatic arthritis and ulcerative colitis were approved in the middle of 2006. Since the study cohort included patients who initiated TNFα antagonists until the end of 2008, we assume a small number of disease misclassifications occurred and did not check for these diagnoses. 2.2.3 The Exposure and Covariates 2.2.3.1 Ascertainment of TNF Antagonist Drug Exposure In analysis of administrative data, dispensing or claims data is used as proxies for drug use. Several possible sources of inaccuracy poses challenges in the analysis: (a) miscoding of the drug itself, strength and dose; (b) compromising for dose adjustment; (c) absent data on free samples and in-hospital drug; and (d) uncertainty that a filled prescription means the drug was actually administered. We discuss these challenges and design features taken to minimize bias in Chapter 3, Section 3.5.3.3, page 80.     40 2.2.3.2 Multiple and Mixed Exposures A major concern when analyzing real life data is multiple and mixed exposure. Some patients experience a serial exposure to more than one drug (switchers), for example. In reviewing previous studies of TNFα antagonists, three approaches for dealing with this concern were identified. The most common approach was to analyze only the first course [79,105-110] (or only the second course [111,112]) of TNFα antagonists. Another approach includes grouping patients based on the first drug administered (the intention-to-treat approach) [63]. In these studies, switching to another drug is considered persistence. This approach is used in studies of therapy persistence (Section 2.1.2.4, page 21). The third approach involves analysis of all drug courses, and patients may be included more than once, depending on the number of courses administered [64, 113-117]. An additional approach, not used in studies of TNFα antagonists is censoring patients at the time of switching to another drug from within the same therapeutic class [118].  In the context of our study, we were interested in isolating patients treated with one drug only. Since previous studies showed differences in persistence between first course and second course, we chose to analyze only the first course of TNFα antagonists.  The second concern was the possible effect of cotreatment (as predictor or as a confounder), especially with regard to methotrexate (MTX). MTX has a role in preventing the formation of antidrug antibody [119-121]. Concomitant MTX therefore may decrease harm [122] and/or improved therapeutic benefit [122-124], however the available evidence is inconclusive. Concomitant MTX has been shown to improve persistence with TNFα antagonists    41 (unpublished data summarized in Appendix A, Section A.3.5, page 261). Consequently, we adjusted the regression model for concomitant MTX.  2.2.3.3 Confounding and Confounding by Indication/Severity One of the major threats to validly in observational studies is confounding, especially confounding by indication (or disease severity). Confounding originates when drug allocation is associated with factors that are also independent predictors the outcome. Confounding by severity occurs when patients with the worst prognosis are assigned the most effective drug [8- 12], and as a result, effectiveness is biased toward the null (no effect). If, for example, drug X is considered more efficacious than drug Y, patients with more severe disease may be treated more frequently with drug X. Disease severity may also affect the outcome, as patients with more severe disease may be less persistent. As a result, drug X that is actually more efficacious is found to be associated with shorter persistence.  Limited or missing information on confounders in analyses of administrative data (such as smoking, alcohol use, the age of menopause, disease severity and diagnostic results) may lead to biased results. We could not evaluate the potential influence of factors relating to patients beliefs, functional level and disease activity, though we used proxies in an attempt to adjust for some of these factors. In addition, biased estimates of possible confounders should be considered in analyses of administrative data. Disease duration, for example, is generally    42 underestimated, because the time of disease onset precedes the time of diagnosis. Several methods have been advocated to minimize confounding, including use of multivariable regression [125], propensity scores15 [125-129] and instrumental variables16 [130-132].  When comparing the effects of two drugs that are prescribed under the assumption of identical effectiveness and safety, as the TNFα antagonists are, predictors of persistence can be considered balanced and minimal bias is expected [133]. In this treatment situation no evidence exists regarding the superiority of either drug (Appendix A, Section A.3.1, page 241) and clinical guidelines recommend treatment with either TNFα antagonist interchangeably (Appendix A, Section A.1.3, page 218). We applied a simple method to adjust for possible confounding - multivariate regression. The main advantages of this approach are that it also allows estimation of the effect of other possible predictors of outcomes, and testing for the confounding effect of each variable separately. A disadvantage of this method is that it allows adjustment only for measured variables.  The variables included in our regression models are presented in Table 3. We sought to adjust for disease severity, accessibility of health services, socioeconomic status and concomitant                                                   15  Propensity score is the “conditional probability of being treated given a certain set of measured covariates. It can be used to control confounding through matching, stratification, regression adjustment or combination of these methods and can be used along with other covariate adjustment” [1]. 16 Instrumental variable is a “method that, under certain assumptions, allows the estimation of causal inference even in the presence of unmeasured confounding for the exposure and effect of interest. An instrumental variable, or instrument, has to meet the following conditions: (1) it is associated with the exposure, (2) it affects the outcome only through the exposure, and (3) it does not share any (uncontrolled) common cause with the outcome” [1].     43 Table 3: The rationale for variables included in the regression model Variable Rationale Demographics  Sex  Females were shown to be less persistent with chronic therapies [134- 136]. In addition, RA was shown to be more severe in females, and progress faster [137]. Age at index date Older age was associated with improved persistence on chronic therapies [135,138,139]. In addition, age may affect the route of administration: younger patients (< 61 years old) were more confident about self-administering drug, and preferred subcutaneous route to intravenous administration [140]. Annual income level (based on MSP subsidy)  The covariate was used as a proxy to both socioeconomic status and out-of-pocket costs [141]. Some studies demonstrated that socioeconomic status influences adherence and persistence with drugs [142,143], but the effect is inconsistent. Out-of-pocket costs were shown to affect TNFα antagonists and other drugs persistence [63,144,145].  Patient’s residence area Residence area influences access to health care facilities and hence influences the selection of the individual TNFα antagonists. It was shown that residence area influenced persistence on chronic drugs [146]. Clinical status  Physicians encounters in the year preceding the index date The number of outpatient encounters was used as a proxy to (a) RA disease severity (b) presence of significant comorbidities, and (c) accessibility of health services. Inpatients admissions in the year preceding the index date The presence of hospital admission was used as a proxy to (a) RA disease severity (b) presence of significant comorbidities, and (c) accessibility of health services. Extra-articular manifestation of RA 17  A proxy measure of disease severity. Comorbidities The presence and severity of comorbidities may influence the selection of the specific drug. In addition comorbidities were shown to affect drug persistence [63,147,148]. Disease duration Disease duration was shown to affect persistence with TNFα antagonists in some studies [114,149] but not other [113,150-152]. Drug therapies  Concomitant MTX  Concomitant and previous use of MTX is strongly associated with improved persistence on TNFα antagonists  in multiple studies [61,64,79,150,151,153-155]. Dispensing claims for NSAIDs A proxy measure of disease activity (pain).    44 Variable Rationale Number of different antirheumatic drugs (synthetic drugs and corticosteroids) dispensed in the three years preceding the index date The number of antirheumatic drugs was used as a proxy of disease severity. In addition, it was shown to be associated with persistence of TNFα antagonists [79,150,156]. Other  Prescriber propensity for treatment discontinuation We hypothesized that different likelihood to discontinuation was associated with different prescribers.  Calendar year at index date Calendar year was included in the multivariable analysis to adjust for secular trends in clinical practice, such as initiating TNFα antagonists in patients with less severe disease and after exposure to a smaller number of antirheumatic drugs[157-160]. This variable also allowed adjustement for later availability of adalimumab, which was marketed since 2004.  MTX, all of which we consider possible confounders. Since we used administrative data, we did not have access to direct measures of disease severity; hence, we used several proxy variables: inpatient and outpatient encounters, disease duration, extra-articular manifestations17, comorbidities, previous number of antirheumatic drugs and nonsteroidal anti-inflammatory drug (NSAID) use (as a proxy for pain). Adjustment for accessibility of health services were performed using geographical area of residence and inpatient and outpatient encounters. A proxy variable for socioeconomic status was the annual deductible level for Fair PharmaCare                                                  17  Extra-articular (outside the joint) manifestations are caused by onvovemnet of tissues and organs outside the joints in the diseas. These include systemic or local manifestations such as weight loss, osteoporosis, lymphadenopathy (enlargments of the lymph nodes), vasculitis (inflammation of the blood vessels) and envolvement of the pleura and pericardium (lining surfaces of the lung and heart, respectively). The list of ICD codes used to identify patients with such manifestations is presented in Appendix B, Table 28, page 3.    45 reimbursement decisions, which is based on the family annual income. Age and sex could also be considered socioeconomic variables.  2.2.3.4 Prescriber-Specific Effects We found that the impact of care-providing physician characteristics on persistence estimates was rarely studied. Generally, comparative persistence was estimated using Cox regression, which included a variety of patient’s characteristics measured at baseline (and rarely were time dependent [150]). Some studies included patient’s psychological factors [152] or adherence/compliance with the drugs [161] as predictors of TNFα antagonist persistence. Until recently, none of the TNFα antagonist studies included adjustment for the care-providing physician or his characteristics18. Rarely, persistence studies in other diseases adjusted for care- providing physician characteristics, such as age [135,162], gender [135,162], number of patients treated by this physician [135,162] or  specialty [104,163-166].  We considered the care-providing physician (and specifically discontinuation habits) to be a source of confounding, and expressed this property as PRESCRIBER PROPENSITY FOR TREATMENT DISCONTINUATION. Since we used administrative data, we could not identify the care-providing physician. We used the prescriber recorded on the first dispensing claim for TNFα antagonist as a proxy of the care-providing physician, but we are aware that                                                  18  A single recent study that examined the effect of the individual physician (prescriber propensity for treatment discontinuation) [48,289] is presented and discussed in Chapter 5, Section 5.5.1.3, page 3.     46 the prescriber may write a prescription based on the recommendation of another physician, with the other physician being the actual care-provider. Generally, two approaches are used to adjust for the prescriber propensity for treatment discontinuation. The first is to include a series of individual Bernoulli variables for each prescriber, to allow an individualization (by prescriber) of the baseline likelihood for discontinuation (Chapter 3, Chapter 4 and Chapter 5). The effect of other predictors in the model is assumed to be constant for all prescribers. The second approach is the use of marginal (Chapter 5) or mixed effect modeling [48] to allow adjustment for possible correlation between patients treated by the same prescriber. In this approach, patients who are treated by the same prescriber are considered more alike than patients treated by different prescribers. In this approach, the effect of the other predictors is averaged.  In our analysis, we included an adjustment for prescriber propensity for treatment discontinuation, using the methods described above. In the analysis described in Chapter 5, we added a new dimension to adjust for the effect of the care-providing physician - PRESCRIBER PREFERENCE FOR THE PRESCRIBED DRUG (one or other TNFα antagonist) as derived from their prescribing patterns. In this analysis, two groups of treatment courses according to the preference level were compared. We assigned value to the preference variable based on previous courses initiated by the same physician in the study cohort (Chapter 5, section 5.3.2, page 132). Additional discussion on modeling of prescribing habits is presented in Chapter 5, Section 5.5.2, page 156.    47 2.2.4 Outcome Variables 2.2.4.1 Ascertainment of Days-Supply An accurate measure of the days-supply for each dispensing is required for ascertainment of the discontinuation date, as discussed in Section 0, page 15 [59]. The British Columbia dispensing data, recorded in pharmacies, as part of the dispensing process, include data on the drug and its strength, the date of dispensing, the quantity dispensed and the number of days- supply. We required an accurate measure of days-supply, which we defined as days until the next dispensing.  Due to the intermittent and complex dosing schedule for TNFα antagonists, we suspected that the recording of days-supply would be inaccurate. We also suspected that inconsistencies could exist in the record of quantity dispensed – weight (in milligrams), volume (in milliliters) or number of vials. Similar concerns were discussed by Wilensky et al in estimating persistence with ophthalmic solutions [167] and by Grymonpre et al regarding nondiscrete dosing forms (such as liquids, creams and ointment) [168]. Initially, we used descriptive statistics to familiarize ourselves with the data. As hypothesized, data on days-supply was problematic and reflected either number of days of administration (for example one day for one dose of infliximab), the number of days until the next dispensing (for example 14 days after the first administration of infliximab) or was set to 30 days based on PharmaCare reimbursement policy [169]. Next, we attempted to calculate the number of days until the next dispensing based on the quantity dispensed. Finally, if required, we used the recorded total cost as the most accurate and reliable field, since this field serves for claim and payment processing.     48 For infliximab (Figure 6, page 49), we based our calculation of days-supply on recorded total cost because both days-supply and dispensed quantity were found to be unreliable. The dispensed quantity was found to have been recorded in a large range (1-1000), with possible overlapping ranges between data recorded as weight, as volume or as number of vials (e.g., quantity of 9 could reflect number of vials or weight). To calculate the days-supply first we estimated the average annual cost per vial, based on dispensing records in which the quantity recorded was unquestionably vials. We expected secular trends in cost and some variation between pharmacies so we averaged the cost per vial in a time-dependent manner: that is estimates depended on the dispensing date. Then we used the annual average cost per vial to roughly estimate the number of vials dispensed in all the records in our data. Next, we calculated days-supply, defined as the number of days until the next expected dispensing event, based on the number of the estimated number of vials dispensed. For patients treated with infliximab, the administered dose is based on body weight. Because we had no access to this data, we used a simple algorithm to approximate the number of days-supply (presented in Figure 6). Finally, to ensure that we did not underestimate the days-supply, the longest period (largest number of days-supply) based on the calculated or the recorded days-supply was selected to be used in the analysis.     49 Figure 6: Algorithm used to establish the number of days-supply for patients treated with infliximab   Map average annual cost per vial, using dispensing records with (a) cost>100 and (b) quantity between 1 and 9 Calculate number of doses based on the number of vials that was calculated in the previous step  Vials Doses 0-2 1 3-5 2 >5 3  For each record in which dispensed quantity is 0 or at least 10, calculate number of vials based on date of dispensing and average cost on dispensing date  (Inconsistency was observed in recorded quantity, which was sometime recorded as mg and not vials)    50  Calculate days-supply   Order of the dispensing Number of doses previously dispensed Calculated number of dose dispensed Calculated days-supply First 0 1 14 First 0 2 48 First 0 3 104 Second 1 1 28 Second 1 2 84 Second 1 3 140 Second 2 Any Doses*56 days 3 rd  and above Any Any Doses*56 days  Calculate order of the dispensing (first, second or 3rd) Select the largest number of days-supply (recorded or calculated)     51 The recorded dispensed quantity for adalimumab was found to be consistent with the number of vials, hence was not manipulated (Error! Not a valid bookmark self-reference., page 51). We used this quantity to calculate the days-supply. For etanercept, the recorded quantity supplied was either vials or weight (in milligrams), hence for large quantities (>99), the number of vials were calculated based on drug strength (25 or 50 mg per vial) (Figure 8, page 52). Again, we used this quantity to calculate the days-supply. Figure 7: Algorithm used to establish the number of days-supply for patients treated with adalimumab  (Recorded quantity was found to be reliable; hence it was used to calculated days-supply) Select the largest number of days-supply (recorded or calculated, minimum 14 days)  Calculate days-supply: dispensed quantity * 14    52 Figure 8: Algorithm used to establish the number of days-supply for patients treated with etanercept   2.2.4.2 Dose Adjustment Dose adjustment, mainly dose escalation, is frequent with the TNFα antagonists as a class and even more frequent in RA patients treated with infliximab (Appendix A, Section A.3.6.2, page 265). In analyses of dispensing claim data, days-supply may be overestimated by calculating dispensed days-supply based on dispensed quantity because dispensed quantity does not reflect that dose escalation is occurring when it occurs. For example, when the actual dose is twice the recommended dose, the patient is dispensed double quantity for a fixed time period (t). To calculate the days-supply, we use the standard recommended dose and do not take into account (Inconsistency was observed in recorded quantity, which was sometime recorded as mg and not vials) For dispensing records where quantity was >99, calculate number of vials (=recorded quantity/drug strength). For other records treated recorded quantity as number of vials Calculate days-supply:  number of vials * drug strength / 50 * 7 days Select the largest number of days-supply (recorded or calculated)      53 the multiplier which represents the dose escalation. Dividing the actual dispensed quantity by the recommended dose produces a value which is twice larger than the actual days-supply (2*t). This value is used in ascertainment of discontinuation and date of discontinuation, and therefore the persistence is overestimated. Since dose escalation is more frequent in infliximab- treated patients then overestimated days-supply are more common with infliximab, causing differential bias. Figure 9: Calculating days-supply in dose escalation may result in persistence overestimation         Patients on recommended dose (R mg/day) Patients on escalated dose, which is n times the recommended dose (n*R mg/day) Dispensing for time interval t    Quantity dispensed: (n*R)*t    Quantity dispensed: R*t Calculated number of days-supply:  (n*R)*t/R = n*t Calculated number of days-supply:  R*t/R = t  Calculating days- supply based on recommended dose (R)    54 Two previous analyses of administrative databases [62,72] evaluated the frequency of dose escalation in RA patients treated with TNFα antagonist. However, in these studies days-supply recordings were assumed to be accurate and were not manipulated as in our analysis. Due to the limitation of our data on dispensed days-supply (Section 2.2.4.1, page 47), it was not possible to identify dose adjustment. 2.2.5 Analytic Approach In the population-based analyses presented in this thesis, comparative persistence is estimated using a multiplicative approach (Cox proportional hazard regression) (Section 2.1.3.1, page 28). This approach is consistent with published research that has also adjusted comparative persistence of TNFα antagonists in RA using the Cox proportional hazard approach [61,110,113,150,151,161,170,171]. Cox regression analyses were also used to compare exposure to other factors, such as age [116,172], concomitant leflunomide [114] or MTX [64,150] and out-of pocket costs [63].  There are two major limitations of this approach in our study. First, it is uncertain whether the proportional hazards assumption holds in studies of pairwise comparative persistence of TNFα antagonists in RA patients. Only one study has reported assessing the proportional hazards assumption [61]. This study demonstrated comparable persistence of adalimumab and etanercept. Nevertheless, Cox regression is ubiquitously used and considered to be a relatively robust modeling approach [173]. A second disadvantage is that we estimate the risk ratio, not     55 the risk difference; hence, statistical significance may not imply clinical significance. Without presenting the absolute difference in persistence, the clinical significance of the results is uncertain.      56 CHAPTER 3:  ESTIMATING COMPARATIVE PERSISTENCE OF TNFα ANTAGONISTS – A POPULATION-BASED STUDY 3.1 Introduction Rheumatoid arthritis (RA) is a relatively common systemic inflammatory disease that causes significant mortality, progressive morbidity and disability and an excessive economic burden. Tumour necrosis factor alpha (TNFα) is a cytokine with an important role in the pathogenesis of this disease [174]. It is bound and inactivated by the TNFα antagonists - infliximab, adalimumab and etanercept. Treatment with a TNFα antagonist is not a cure [44] but rather reduces the symptoms of inflammation, possibly induces remission and possible prevents long- term complications. The individual TNFα antagonists have different pharmacological properties which may lead to substantive differences in therapeutic benefits and harms (Appendix A, Section A.2.4, page 229, [39-43]); however, their relative benefit-harm balance is not well studied and there are no randomized clinical trials (RCTs) that directly compare therapeutic outcomes of all three or pairs of the drugs. Knowledge about relative benefit-harm balance is important to patients, clinicians, policy makers and researchers to optimize clinical and policy decisions and to test the validity of studies. Drug persistence is considered an indirect simple approach that balances overall long-term benefit (effectiveness) and harm (safety, adverse events) of drug therapy, particularly in patients with chronic noncurable diseases [45]. Beside benefit and harm, it integrates the    57 effects of factors not primarily related to drug use, such as cost. Comparative persistence, which is a relative measure, is intuitively considered more accurate than absolute measurement (persistence with a drug) since the effect of systematic errors (e.g., exposure and outcome classification biases) is diminished. We have estimated comparative persistence with the TNFα antagonists in RA patients in the Canadian province of British Columbia. 3.2 Objectives and Hypothesis  The goal of the current study was to estimate pairwise comparative persistence with TNFα antagonists infliximab, adalimumab and etanercept in RA patients. The study hypothesis was that variations in pharmacological characteristics of the individual TNFα antagonist drugs caused statistically and clinically significant differences in persistence. 3.3 Patients and Methods 3.3.1 Data Source and Study Cohort  We analyzed a cohort of British Columbia residents who received a first course of a TNFα antagonist between March 2001 and December 2008, and had also been diagnosed with RA. Follow-up data were available until December 31, 2009. Patients were identified using four British Columbia Ministry of Health administrative databases: PharmaNet (prescription    58 dispensing data), Medical service Plan (MSP) registration information (demographic data), MSP Payment Information (fee-for-service payments to physicians and alternative providers), and the Discharge Abstract Database (hospital separations). The databases were linked using a study-unique anonymized identifier. The study protocol was approved by the Clinical Research Ethics Board of the University of British Columbia.  The study cohort was identified based on the algorithm presented in Appendix B, Figure 23, page 270. Patients were included based on one or more recorded dispensing claims for infliximab, adalimumab or etanercept between March 2001 and December 31, 2008. The index date was defined as the date of the first dispensing of a TNFα antagonist after a run-in period of at least 36 months of follow up without TNFα antagonist dispensing. RA patients were selected using similar, though not identical, criteria to previous studies in British Columbia RA patients [101-103]. RA patients were identified based on two outpatient visits with a diagnosis code of RA at least 60 days apart, or one hospitalization with a recorded discharge diagnosis of RA in three years prior to the index date. A list of International Classification of Diseases (ICD) codes that were used in the study is presented in Appendix B, Table 28, page 271. Diagnostic codes were recorded in the databases for billing purposes and were based on clinical judgment of the care-providing physician. Additionally, patients were required to have continuous provincial MSP coverage three years before the index date. A gap shorter than 30 days was not considered to be an interruption. Patients were excluded if they were previously treated with anakinra, rituximab or abatacept, if sex or date of birth were missing, if they had a concurrent diagnosis of Crohn’s disease (based on at least one outpatient or inpatient diagnosis code in the three years preceding the index date), or if they were younger than 18 years old at the index date (to remove patients with juvenile RA).    59 3.3.2 Exposure and Covariates The initial TNFα antagonist exposure was identified for each patient based on their pharmacy dispensing records. We limited the analysis to a single exposure categorical variable for each patient.  The choice of a specific TNFα antagonist drug could be associated with disease severity or additional factors that influence drug persistence. Therefore, we also included demographic and clinical status variables in the model to adjust for possible confounding bias. The demographic variables included sex, age and a proxy variable for socio-economic status (the annual deductible for prescription cost, which was based on family annual income [141]). Clinical status variables included the number of inpatient and outpatient encounters in the year prior to initiating treatment, the duration of disease (defined as time from the first recording of diagnostic code of RA in the data to the index) and the presence and severity of comorbidities. We used Quan’s ICD-9-CM and ICD-10 coding algorithm for administrative databases [175], excluding rheumatic diseases, to determine the Charlson comorbidity score [176] for each patient in the cohort during the year preceding the index date. We also included three variables for other RA drugs: concomitant methotrexate (MTX) based on at least one dispensing in the 200 days prior to index19 (1=yes, 0=no) and dispensing of any nonsteroidal anti-inflammatory                                                  19  The period was selected based on mean plus two standard deviations of between-dispensing intervals of MTX in the study cohort    60 drug (NSAID) in the year preceding the index date (1=yes, 0=no). The third variable, the number of previous antirheumatic drugs was based on the dispensing of the following nine drugs within three years prior to the index date: MTX, hydroxychloroquine, sulfasalazine , azathioprine , cyclosporine , minocycline, penicillamine, sodium aurothiomalate, prednisone and intra-articular triamcinolone/methylprednisolone acetate. We included a categorical variable for the calendar year of the index date, which allowed adjustment for possible secular trends in clinical practice [157-160] and late availability of adalimumab. Finally, we included several Bernoulli variables for the 12 highest volume prescribers (physicians who initiated > 70 courses in the study cohort), to adjust for the prescriber propensity for treatment discontinuation. The list of covariates included in the models and their forms are presented in Table 4. The rationale for considering the covariates is presented in Chapter 2, Section 2.2.3.3, page 41.  3.3.1 Outcomes The two outcomes variables were (a) a continuous persistence variable and (b) a binary censor variable. Drug persistence in years was calculated as the difference between the date of drug discontinuation and the index date. Drug discontinuation was ascertained by either switching to another antirheumatic monoclonal antibody or immunosuppressant drug (infliximab, adalimumab, etanercept, anakinra, rituximab, abatacept, certolizumab and golimumab), or elapsing of a drug-free interval of 180 days. A drug-free interval was defined as a period without additional dispensing of the same pharmaceutical component after the days-supply of     61 Table 4: List of covariates included in the final models Variable Period considered in assigning the value of the variable Description Demographics   Sex  Bernoulli, reference=female Age  Index date Four mutually exclusive categories: 18-29, 30- 69 (reference), 70-79 and ≥F80 years old Annual deductible for prescription cost Index date Six mutually exclusive categories: $0 (reference), 1-500, 501-2250, >2251, other plan, no plan Clinical status   The number of previous outpatient encounters  One year preceding the index date Continuous The number of previous hospital admissions  One year preceding the index date Four mutually exclusive categories: none (reference), 1, 2, >2 The presence and severity of comorbidities  One year preceding the index date Five mutually exclusive categories: Charlson comorbidity score of 0 (reference), 1, 2, 3, >3 RA disease duration From the beginning of the data until the index date Four mutually exclusive categories: 0-3 years (reference), 3-8, 8-12, >12 Drug therapies   Concomitant MTX 200 days preceding the index date Bernoulli, reference=no NSAIDs use One year preceding the index date Bernoulli, reference=no The number of previous antirheumatic drugs  Three years preceding the index date Four mutually exclusive categories: none, 1-2, 3-6 (reference) , >6 Other   Calendar year at index date  Index date Eight yearly categories were included for the years 2001-2008 (reference = year 2001). Prescriber propensity for treatment discontinuation Index date 12 Bernoulli variables for individual prescribers, reference=low-volume prescribers (<70 courses in the study cohort)    62 the latest dispensing was exhausted (Chapter 2, Figure 5, page 22). The date of drug discontinuation was set to the end of days-supply or the date of the first dispensing of a different antirheumatic monoclonal antibody or immunosuppressant drug, whichever was earliest. Patients were considered censored if they were continuously treated with the first TNFα antagonist on December 31, 2009 (end of  follow-up period) or if they discontinued the drug during an interruption of more than six days in the provincial MSP coverage. Death and emigration from the province are the most common causes of interruptions in the provincial MSP coverage. For details on assigning the values of the outcome variables refer to Appendix B, Figure 24, page 271. Original data on the number of days-supply recorded in PharmaNet were unreliable (Chapter 2, Section 2.2.4.1, page 47); hence we also calculated the expected number of days-supply based on estimated vials dispensed, which was imputed using the cost field in the dispensing record (Chapter 2, Figure 6, Figure 7 and Figure 8, from page 49). We used the longest duration of days-supply, recorded or calculated, to determine both the length of drug-free interval and the date of drug discontinuation. For discussion on calculating the number of days- supply refer to Chapter 2, Section 2.2.4.1, page 47.  3.3.2 Statistical Methods  Summary statistics of baseline characteristics were compiled across the three drug groups. We assumed normality of the continuous variables compatible with large sample size (>2500) and the central limit theorem. The significance of differences between the drug groups was    63 assessed with one-way analysis of variance (ANOVA) F-test for continuous variables and the Pearson’s Chi-square test for categorical variables. All statistical tests were two-sided.  The product-limit method and the log rank test were used to estimate and compare levels of drug persistence. Pairwise comparisons between the three drug groups were planned a priori but were considered only when the overall comparison indicated a significant difference between TNFα antagonists (a two-tailed p-value < 0.05). The crude and adjusted hazard ratios for drug discontinuation were estimated using Cox proportional hazards regression. Three pairwise comparisons were presented: infliximab versus adalimumab, infliximab versus etanercept and adalimumab versus etanercept. Drug-sex and drug-age interactions were examined but were included in the final model only if the overall significance was < 0.1. The proportional hazards assumption was assessed by testing for the significance of interactions between the variables and time or the natural logarithm of time, as well as by plotting the scaled Schoenfeld residuals [177]. The validity of the linearity assumption of continuous variables was assessed by log likelihood tests comparing models that included either categorical variable or polynomial variables. If nonlinearity was detected, the variable was categorized. We also presented the adjusted hazard ratios for all variables included in the model. In addition, the confounding effect was presented for covariates that increased or decreased the crude adjusted hazard ratios (or their 95% confidence limits) by at least five percentages. All calculations were performed using the SAS software package [178].     64 3.4 Results Between March 2001 and December 2009, a total of 7,212 British Columbia residents were treated with TNFα antagonists. The study cohort included 2,923 RA patients who initiated TNFα antagonists by December 31, 2008 and met all selection criteria. Of those, 2104 (72%) were females and age ranged from 18-92 years (mean± standard deviation [SD], 55.9±13.6 years). Patients treated with adalimumab were older and had more previous hospital admissions and lower prevalence of concomitant MTX and dispensing claims of NSAIDs compared to patients on infliximab and etanercept. The baseline characteristics across the three drug groups are presented in Table 5.  3.4.1 Comparative Persistence With TNFα Antagonists The Kaplan Meier curves for persistence are presented in Figure 10, page 67. The median persistence with infliximab was 3.68 years (95% confidence interval (CI) 2.88-4.95), with adalimumab – 3.33 (95% CI 2.63-4.10) and with etanercept 3.78 (3.31-4.32). Insignificant differences in persistence were demonstrated between the three individual drugs (log rank test p-value of 0.23). In the multivariable analysis, the estimated adjusted hazard ratios were comparable for the three TNFα antagonists’ contrasts (Table 6, page 68). Drug-sex or drug-age interactions were not significant.   65 Table 5: Baseline characteristics  Variable All patients (N=2923) Infliximab (N=620, 21%) Adalimumab (N=474, 16%) Etanercept  (N=1829, 63%) P-value for comparison Females, n (%)  2104 (72%) 438 (71%) 344 (73%) 1322 (72%) NSS Age at index (years) median (range) 56 (18-92) 56 (18-87) 58 (22-91) 56 (18-92) 0.002 20   Annual deductible for prescription cost n (%)        None 345 (12%) 49 (8%) 83 (18%) 213 (12%) <0.0001   $1-500 252 (9%) 34 (5%) 56 (12%) 162 (9%) 0.001   $501-2250 570 (19%) 96 (15%) 118 (25%) 356 (19%) 0.0005   >$2250 255 (9%) 40 (6%) 46 (10%) 169 (9%) NSS The number of previous outpatient encounters median (range) 32 (2-158) 33 (3-158) 31 (2-112) 32 (3-136) NSS  The number of previous hospital admissions median (range) 0 (0-8) 0 (0-6) 0 (0-5) 0 (0-8) 0.01 21   The presence and severity of comorbidities (Charlson comorbidity score)  median (range) 0 (0-8) 0 (0-6) 0 (0-8) 0 (0-7) NSS                                                   20  Infliximab versus adalimumab p=0.0007; adalimumab versus etanercept p=0.0023 21  Infliximab versus adalimumab 0.0115; adalimumab versus etanercept 0.0025    66 Variable All patients (N=2923) Infliximab (N=620, 21%) Adalimumab (N=474, 16%) Etanercept  (N=1829, 63%) P-value for comparison RA disease duration (years) median (range) 8.3 (0-17.9) 9.1 (0.1-17.9) 7.7 (0.3-17.9) 8.2 (0-17.8) NSS  Concomitant MTX n (%) 1596 (55%) 417 (67%) 277 (59%) 902 (49%) < 0.0001 22   Dispensing claims of NSAIDs n (%) 1539 (53%) 332 (54%) 224 (47%) 983 (54%) 0.0367  The number of previous synthetic antirheumatic drugs median (range) 4 (0-9) 4 (0-8) 4 (0-8) 4 (0-9) NSS Calendar year at index date median (range) 2005 (2001-2008) 2003 (2001-2008) 2007 (2004-2008) 2005 (2001-2008) < 0.0001 23   %- percent; $-Canadian dollars; MTX-methotrexate; N- number of patients in this drug group; n-number of patients; NSAID- nonsteroidal anti- inflammatory drug; NSS – not statistically significant                                                  22  Infliximab versus adalimumab p-value= 0.0027; infliximab versus etanercept <0.0001; adalimumab versus etanercept 0.0004 23  Adalimumab versus etanercept p-value=0.01    67  Figure 10: Drug persistence curves    68 Table 6: Hazard ratios for discontinuing TNFα antagonists  Crude hazard ratio  (95% CI), p-value Adjusted hazard ratio  (95% CI), p-value Infliximab versus etanercept 0.92 (0.81-1.04), p=0.19 0.98 (0.85-1.13), 0.76 Infliximab versus adalimumab 0.87 (0.73-1.03), p=0.10 0.95 (0.78-1.15), 0.57 Adalimumab versus etanercept 1.06 (0.92-1.23), p=0.42 1.04 (0.88-1.22), 0.68 CI – confidence interval 3.4.2 Predictors of Drug Discontinuation Female sex and younger or older age (<30 years or > 70 years) were predictors of discontinuing the TNFα antagonists (Table 7), while the deductible cost had no effect on the risk for discontinuation. The hazard for discontinuation was increased in patients with admission to hospital in the year preceding the index date, but the number of outpatient encounters, comorbidity and disease duration had no effect. Concomitant MTX, dispensing claims of NSAIDs and smaller number of previous synthetic antirheumatic drugs (<7) were all associated with decreased hazard for discontinuation. Lastly, prescriber propensities to treatment discontinuation were significantly associated with the hazard for discontinuation.    69 Table 7: Statistical significance of covariates Variable P value (chi- square log likelihood test) 24  Category Hazard ratio (95% CI), p-value for type 3 Wald chi-square statistic Sex (reference female)   0.76 (0.68-0.86), <0.0001 Age (reference 30-70 years)           <0.0001 18-29 years 1.53 (1.20-1.96), 0.0007 70-79 years 1.32 (1.14-1.54), 0.0003 ≥80 years 1.83 (1.39- 2.41), <0.0001 Annual deductible for prescription cost (reference $0) 0.80 $1-500 0.94 (0.74-1.18) $501-2250 1.01 (0.83-1.23) >$2250 0.87 (0.67-1.12) No plan 1.00 (0.82-1.22) Other plan (not income-based) 1.02 (0.83-1.25) The number of previous outpatient encounters  Increase of one encounter 1.003 (1.000-1.006), 0.051                     The number of previous hospital admissions (reference 0) 0.006 1 admission 1.26 (1.09-1.47), 0.003 2 admissions 1.10 (0.86-1.42) >2 admissions 1.51 (1.09-2.10), 0.02 The presence and severity of comorbidities (reference Charlson comorbidity score of 0) 0.23 Score of 1 1.01 (0.87-1.17)  Score of 2 1.12 (0.89-1.42) Score of 3 1.24 (0.86-1.79) Score >3 0.62 (0.35-1.10) RA disease duration (reference 8-12 years) 0.07 0-2 years 1.19 (0.997-1.43) 2-4 years 1.08 (0.91-1.30)  4-8 years 0.94 (0.81-1.10) >12 years 0.95 (0.80-1.11)                                                  24  For categorical variables with more than two categories    70 Variable P value (chi- square log likelihood test) 24  Category Hazard ratio (95% CI), p-value for type 3 Wald chi-square statistic Concomitant MTX (reference no)  Yes 0.79 (0.71-0.88), <0.0001 Dispensing claims of NSAIDs (reference no)  Yes 0.89 (0.80-0.99), 0.03  The number previous synthetic antirheumatic drugs (reference 3-6 drugs) 0.03 None 1.38 (0.99-1.92) 1-2 drugs 1.04 (0.90-1.20) 7-9 drugs 1.40 (1.09-1.81), 0.009 Calendar year at index date (reference 2005) 0.08 2001 0.79 (0.63-0.997), 0.047 2002 0.78 (0.62-0.97), 0.03 2003 0.71 (0.57-0.88)  2004 0.84 (0.70-1.02) 2006 0.89 (0.73-1.08) 2007 0.84 (0.69-1.03) 2008 0.76 (0.60-0.96), 0.02 Prescriber propensity for treatment discontinuation 25  (reference prescribers with <70 courses) <0.0001 Prescriber 0.57 (0.42-0.76) , 0.0001    Prescriber 0.67 (0.47-0.98), 0.04 Prescriber 1.38 (1.03-1.84), 0.03 Prescriber 1.57 (1.26-1.96), <0.0001 $- Canadian dollars; CI- confidence interval; MTX- methotrexate; NSAIDs- nonsteroidal anti-inflammatory drugs; RA- rheumatoid arthritis                                                  25  Only four prescribers with significant hazard ratios are presented    71 3.4.3 Confounding Concomitant treatment with MTX had a confounding effect on the hazard ratio of infliximab versus etanercept. Confounding was defined as an increase or decrease of the crude hazard ratios (or their confidence intervals) by at least five percentage points. Prescriber propensity for treatment discontinuation had a confounding effect on hazard ratios of adalimumab (infliximab versus adalimumab and adalimumab versus etanercept). None of the remaining covariates had a large influence on the magnitudes of the estimated hazard ratios (Table 8). Table 8: Confounding effects  Model Hazard ratio (95% CI)  Infliximab versus etanercept  Infliximab versus adalimumab Adalimumab versus etanercept No adjustment 0.92 (0.81-1.04) 0.87 (0.73-1.03) 1.06 (0.92-1.23)  Adjustment for concomitant MTX  0.98 (0.86-1.11)   Adjustment for prescribers  0.90 (0.76-1.08) 1.01 (0.87-1.17) Data is presented for covariates that changed the magnitude of point estimate (or confidence limits) of the hazard ratios by at least 5% only  MTX- methotrexate    72 3.5 Discussion 3.5.1 Similar Persistence With Infliximab, Adalimumab and Etanercept  The hypothesis that different TNFα antagonists have different persistence in RA patients was not supported by the results of an analysis of British Columbia data. There are several explanations for the finding that treatment with infliximab, adalimumab and etanercept were associated with similar persistence in RA patients. One explanation is that drug persistence is not an accurate measure of benefit-harm balance, due to two reasons. First, while most RA patients discontinue TNFα antagonists for lack of benefit or perceived harm (Appendix A, Section 0, page 255), they also discontinue for a variety of other reasons such as remission [170,179], planned pregnancy [108,170,180], financial constraint [68] and patient or prescriber preference for subcutaneous administered drugs [181]. Imbalance between the drug groups with regard to any of these reasons could introduce bias. For example, discontinuation due to preference for subcutaneous administered drugs was applicable only for patients who were treated with infliximab. As a result, infliximab discontinuation in these patients was not associated with impaired benefit or increased harm. Second, the intravenous administration of infliximab requires regular physician follow-up that has been shown to encourage adherence and persistence to drug therapy in a variety of diseases [162,182-185]. As a result, persistence with infliximab may not  reflect solitarily drug benfit-harm balance. Therefore, in studies comparing persistence with infliximab to the other drugs, persistence may not be a valid measure of differences in benfit-harm balance.    73 Assuming, however, that persistence is an indirect and valid measure that balances benefit and harm leads to an examination of alternate explanation for the findings. Comparable effectiveness amongst the three TNFα antagonists may result from conflicting effects of the pharmacological characteristics and the route of administration of the three drugs. We demonstrate the conflicting effects using the example of the characteristics of infliximab. Several pharmacological characteristics of infliximab may decrease its effectiveness or increases its harmful effects compared to etanercept and adalimumab. Higher immunogenicity of infliximab compared to adalimumab and etanercept [186,187] may lead to increased risk of harm, mainly as infusion reaction. Additionally, the steady-state concentration-time profiles of infliximab is less smooth then with the other two drugs, with wide fluctuations in serum concentration [42,188]. This increases the risks of exceeding the maximum tolerated concentration leading to harm, or of reaching suboptimal concentrations leading to lack of therapeutic benefit. Lastly, interferon gamma production is inhibited by infliximab but not etanercept [189] and may result in excessive risk for tuberculosis.  On the other hand, several characteristics may be associated with improved effectiveness with infliximab compared to etanercept t. First, infliximab can causes apoptosis (cell death) of activated T cells and circulating monocytes [189], which is a potentially desirable property in chronic inflammation. Additionally, etanercept is the only drug that binds to lymphotoxin alpha, with an affinity that is similar or even higher than to soluble TNFα. This may results in unneutralizing the soluble     74 TNFα [188]. Lymphotoxin alpha was suggested to attenuate the function of TNFα [190] and therefore blocking both lymphotoxin alpha and TNFα activity resulting in higher activity of residual TNFα and decreased benefit.  Comparable persistence with the three drugs was previously demonstrated in studies from Europe [108,113,171] and the United State [72]. Additional studies compared two of the drugs and found similar persistence of infliximab and etanercept [191] or adalimumab and etanercept [61,192]. On the other hand, multiple studies have demonstrated differences in persistence [62,106,110,150,151,161,170,193,194]. An in-depth exploration of the reasons for the differences between the studies was beyond the scope of this study. Implications Accepting the assumption that drug persistence is a valid indirect measure that integrates both therapeutic benefit and harm, we can infer that the benefit-harm balance of the three TNFα antagonists is comparable. In this case, drug selection should be based on other factors such as patient preference for route of administration and cost. Policy of drug reimbursements plans should be recommended based on costs, availability etc. Additional Research Discrepancies between published studies in the significance and direction of estimates of comparative persistence with the three TNFα antagonists require further research. Consequently, we conducted a systematic review and meta-analysis designed to pool the estimates and explore possible factors that contribute these discrepancies (unpublished study). In this study, we found substantial between-studies heterogeneity in three pairwise comparisons of persistence with TNFα antagonists. We demonstrated that in the     75 research published to date, heterogeneity arises mainly from the use of different types of data sources and suggested that it could also arise from the use of different methodological approaches to ascertain discontinuation. Two main directions of future research are proposed. First, based on the abovementioned unpublished systematic review and meta-analysis ,we hypnothesize that the heterogeneity in comparative persistence estimates that was associated with type of data source could be caused by the different methods used in the ascertainment of discontinuation. This hypothesis was examined in Chapter 4. Specifically, since we only had access to administrative data, we applied several different algorithms to ascertain discontinuation, and then examined their influence on estimates of comparative persistence. Second, we suggest research to direct support of the association between therapeutic benefit and harm of TNFα antagonists, such as decrease in disease severity during treatment, and following discontinuation. Access to time- series clinical data on changes in disease activity and therapeutic harm would be required.  3.5.2 Possible Confounders In this analysis of British Columbia data, we demonstrate that including variables for concomitant MTX and prescriber propensity for treatment discontinuation changed the magnitude of the crude hazard ratios (or their confidence interval) by at least five percentages. These factors should therefore be considered to be confounders and be included in future     76 studies of comparative persistence. The effect of prescriber propensity for treatment discontinuation was consistently ignored in studies of persistence with TNFα antagonists in RA and in other clinical situations26.  We consider several explanations for differences in discontinuation risk in patients treated by different physicians (prescriber propensity for discontinuation). First, differences in clinical practice and prescribing habits because of differences in education or adherence to different guidelines (Appendix A, Section A.3.2, page 251) may lead to differences in the likelihood of discontinuing TNFα antagonists. Second, patients treated by the same prescriber are probably more similar to each other than patients treated by different prescribers are (correlation). We discuss analytic methods used to adjust for the risk differences (prescriber propensity for discontinuation) in Chapter 5, Section 5.5.2, page 156. Implication Studies of persistence with TNFα antagonists in RA patients should, in our opinion, always include the following variables in the final model, regardless of the predictor of interest: concomitant MTX and prescribers.                                                    26  A recent study of TNFα antagonist discontinuation in RA patients (published after the completion of the thesis research) included clustering by physician to adjust for the possible effect of the individual physician (the degree to which the physician favors discontinuation) [48] . The study is discussed in Chapter 5, Section 5.5.1.3, page 3.     77 Future Research The ways that the prescriber could influence drug persistence require further exploration. Testing for possible correlation between patients treated by the same prescriber could be conducted by comparing the fit of models for clustered data, such as mixed effect or marginal model and models for noncorrelated data. In Chapter 5, we present a study conducted to test for the hypothesis that prescriber preference for the prescribed drug influences the persistence.  3.5.3 Strengths and Weaknesses of the Study One of the main strengths of our population-based study our data source of provincial administrative claims for drugs and medical encounters. The universal nature of the Canadian health care system should have minimized selection bias thus increased the external validity of our results. Access to complete information on all dispensed prescription drugs and multiple sources of data (physician encounters, hospital separation and pharmacies) contributed to the usefulness of the data, and made the cohort an excellent resource for pharmacoepidemiologic research.  The study was characterized by large sample size (N=2923), which provided ample power as well as relatively long durations of follow up in our study patients (roughly nine years from the enrollment of the first patient). Furthermore, the British Columbia Ministry of Health uses a largely systematic and standardized approach to data collection, which ensures the better representativeness of real life treatment patterns and drug-taking behaviour [133] and therefore generalizability [93]. Selection bias should have been minimized because our population-based    78 cohort included all patients regardless of their disease severity, duration of disease, comorbidities or concomitant treatments [93]. Administrative data do not suffer from recall bias27 [133]. Analysis of administrative data is also inexpensive, unobtrusive and quick. The results are often referred to as ‘signals’ that warrant further research using clinical data. Measurements of drug utilization (persistence and compliance) using administrative data have been validated through patient surveys [168,195,196], medical charts [197], pill counting [168,195] electronic devices [196] and simulation [73]. However, each method has its limitations and a gold standard method to measure persistence and compliance does not exist [198].  This study was designed to minimize the effect of potential sources of bias caused by inaccuracies in data recordings (diagnosis and drug refills), possible confounding effects and possible limited generalizability – all are sources of bias that should be considered in an observational study.                                                   27  Recall bias is a “systematic error due to differences in accuracy or completeness of recall to memory of past events or experiences” [1]     79 3.5.3.1 Inaccuracy of Diagnosis Recording  Disease diagnoses are coded and recorded in British Columbia databases for payment purpose; hence, accuracy might be compromised or biased. There are multiple opportunities for inaccuracies to arise, including misdiagnosis, miscoding of diagnosis28 and undercoding (failure to record complete information), especially of secondary diagnoses - all these sources of bias may be random or systematic. In addition, differences in accuracy of codes between the care-providers (e.g., different physicians) and variation in coding precision in different diseases are to be expected. To minimize the effect of these sources of inaccuracy, we designed a disease algorithm requires at least two outpatient encounters with RA diagnosis (Chapter 2, Section 2.2.2, page 35). 3.5.3.2 Diagnostic Shift  Diagnosis shift is a change in the frequency of the diagnosis coding without actual change in the disease frequency, and can be caused by change in clinical practice, payer guidelines or in coding systems. Coding shifts have been demonstrated in several diseases [199-206], but not in RA or other rheumatologic diseases and                                                   28  The ICD-9 and ICD-10 coding systems do not provide diagnostic criteria to diseases, nor do these systems enforce the use of specific diagnostic criteria. Social, psychological and functional aspects of the health status are often not captured by the coding system [480]. In addition, in some clinical situation several codes may be appropriate and the coding may be motivated by preference of the payer [92]    80 therefore we did not consider it a possible source of bias. Additionally, we included a categorical variable to adjust for the calends year of treatment initiation, that could adjust for diagnostic shift, if exists. 3.5.3.3 Drug Exposure  In analysis of administrative data, dispensing or claim data is used as proxies to drug use. Several sources of inaccuracy should be considered. 1. Miscoding of the drug itself, strength, dose and dispensed days-supply would influence the ascertainment of drug exposure as well as the estimate of persistence.  2. Inaccuracy in dispensed days-supply - The number of days-supply recorded is used to ascertain drug use, as well as drug discontinuation. Generally, the number of days-supply reported by the pharmacists should be validated, especially when an intermittent dosing schedule is in use. Systematic errors in recording days-supply might be as a result of the refill policy of the drug-coverage plan, different interpretations of days-supply (Chapter 2, Section 2.2.4.1, page 47), dose titration, unknown actual use, and drugs used only as needed [93]. In this analysis using British Columbia data, we found that diverse approaches were used by pharmacists to record days-supply; hence calculation of actual days-supply was required (Chapter 2, Section 2.2.4.1, page 47). By using a conservative estimate of days-supply (the longest among several possible approaches to estimations), we might have underestimated the actual length of drug-free intervals and therefore underestimated the    81 number of patients defined as discontinuers. It is possible that inaccuracies in the calculated number of days-supply biased the ascertainment of  drug discontinuation. Nevertheless, since the data for the three drugs were treated similarly, we believe that the hazard ratios were not systematically biased in this regard. We calculated the actual number of days-supply based on dispensing cost. We regarded cost the most accurate data, because it was necessary for billing and payment processes. Estimation of days-supply for patients treated with infliximab was especially challenging. Since infliximab dosing is based on weight [207] and the patient weight was inaccessible, we used an algorithm based on number of vials and assumption of weight < 70 Kg. This could lead to overestimation of days-supply in heavier patients. Consequently, the persistence with infliximab was potentially overestimated, which could bias the estimates of comparative persistence of infliximab versus adalimumab or etanercept.  3. Dose adjustment influences both the calculated number of days-supply and the date of drug discontinuation but may be challenging to ascertain. Two analyses of administrative databases [62,72] previously evaluated the frequency of dose escalation in RA patients treated with TNFα antagonist, but in these studies days-supply recordings were regarded as accurate and were not modified as in our analysis. Based on our data and the requirement to estimate days-supply based on cost, we could not identify patients with dose adjustment. Therefore, we may have overestimated the calculated days-supply for patients who experienced dose escalation (Chapter 2, Section 2.2.4.2, page 52). Dose escalating has been found to be more common in patients treated with infliximab (Appendix A, Section A.3.6.2, page 265), and therefore the estimated comparative persistence of  infliximab    82 versus adalimumab or etanercept were possibly biased toward overestimated persistence with infliximab. Further discussion is presented in Chapter 2, Section 2.2.4.2, page 52. 4. Absent data on free samples and in-hospital drugs may cause selection bias, as not all patients treated are included, especially those with shorter persistence, and also lead to an underestimate of persistence in included patients.  5. Even if dispensing data is accurate, though filling a prescription is generally regarded as an indication of the patients willingness to persist, there is no certainty that a filled prescription means the drug was actually administered [208]. As a result, drug persistence may be overestimated. 3.5.3.4 Outcome Misclassification We used a long drug-free interval to ascertain discontinuation; this design minimized outcome misclassification (i.e. persistent patients are considered discontinuers). A long drug-free interval minimized the number of patients considered discontinuers for a temporary interruption in persistence. We claim that temporary interruptions should be considered noncompliance, not discontinuation. By using short drug-free intervals, previous studies captured a complex measure of both persistence and compliance and therefore the end point in these studies should be considered nonpersistence and not discontinuation [77]. Further discussion is presented in Chapter 2, Section 2.1.2.4, page 21 and Chapter 4, Section 4.5.1, page109.     83 3.5.3.5 Confounding We designed the study to address potential confounding, especially confounding by indication and/or severity – a major concern in a non-randomized observational study. To diminish possible confounding, a threat to internal validity, we used multivariable regression methods. To adjust to possible differences in disease activity, we used the following proxies in multivariable regression:  (a) sex (RA is more severe in females [137]), (b) the number of previous outpatient encounters and the number of previous hospital admissions, (c) disease duration, (d) dispensing records of NSAIDs (as a proxy for pain) 29 and (e) number of previous synthetic antirheumatic drugs the patient was exposed to29. Further discussion on confounding is presented in Chapter 2, Section 2.2.3.3, page 41. In this study, we identified imbalances between the drug groups with regard to age distribution, annual deductible for prescription costs, the number of previous hospital admissions, concomitant MTX, dispensing claims for NSAIDs and calendar year at index. We found that age, the number of previous hospital admissions, concomitant MTX and dispensing claims for NSAIDs were also predictors of persistence and therefore were likely to confound the results.                                                  29  Dispensing records of NSAID and number of previous antirheumatic drugs were identified in data fron the year preceding treatment initiation.    84 3.5.3.6 Absence of Clinical Data The study was designed to compensate to the absence of access to clinical data by using proxy measures for drug exposure and clinical variables (discussed above). In addition, while persistence in years and the rate of discontinuation could be estimated, the reasons for discontinuations, specifically the contribution of decreased benefit or perceived harm cannot be determined using data from administrative databases. However, since most guidelines recommend at least three months of treatment with TNFα antagonists before ascertainment of lack of benefit, a reasonable assumption is that discontinuation within the first three months of treatment was caused by harmful events.  3.5.3.7 Generalizability Threats to generalizability could potentially be caused by coverage exemptions -   First Nation and federal employee data was absent from the British Columbia data available to us. In addition, patients who left the province would be lost to follow up and this could potentially affect the outcome estimate. This study is characterized by population-based data and a study design that attempted to minimize the effect of multiple sources of bias. Therefore the estimates of comparative persistence are of high quality and be easily generalized.     85 3.6 Conclusions  The study results indicate that persistence with infliximab, adalimumab and etanercept is similar in RA patients. Accepting the assumption that persistence accurately reflects the benfit-harm balance of the therpy,  the drugs’ benefit-harm balance could be considered equivalent. Accordingly, clinical and policy decisions for treatment with TNFα antagonist should be based on convenience and cost.   The study was designed to minimize the effect of multiple possible sources of bias. Among other features, we used a disease algorithm with a high degree of specificity, ascertained discontinuation using a long drug-free interval, corrected recorded days-supply to minimize outcome misclassification; and adjusted for multiple possible confounders, including prescriber propensity for treatment discontinuation.   Enforcement of an improved and uniform policy for recording days-supply in the British Columbia PharmaNet would improve the accuracy of measurements of drug persistence.     86 CHAPTER 4:  SENSITIVITY OF COMPARATIVE PERSISTENCE ESTIMATES TO THE METHODS USED TO ASCERTAIN DISCONTINUATION  4.1 Background We study comparative persistence with tumour necrosis factor alpha (TNFα) antagonists in patients with rheumatoid arthritis (RA). Comparative drug persistence was suggested as an indirect measure with which to compare overall benefit-harm balance  of competing drugs in patients with noncurable diseases [45]. This measurement approach is based on the assumption that a drug is continued in these patients provided that it has therapeutic benefit and only mild harmful effects that the benefit overweights [45,209]. Knowledge on the relative benfit-harm balance of the three TNFα antagonists – infliximab, adalimumab and etanercept – is limited due to the absence of head-to-head RCTs and unrepresentativeness of RCT participants (Appendix A, Section A.3.1, page 241 [210-212]). Studies have shown that the therapeutic benefit of TNFα antagonists is exaggerated in RCTs compared to routine practice as a result of unrepresentativeness of RCT participants (Appendix A, Section A.3.1.2, page 246, [210-212]).  Drug persistence can be estimated using clinical data, administrative (claim) data or patient surveys; however, only data from the first two sources were utilized in studies of TNFα antagonists in RA (unpublished original systematic review and meta-analysis [49,50]. In an original unpublished systematic review and meta-analysis of comparative persistence of TNFα antagonists in RA patients, we found heterogeneity in the magnitude, significance and    87 direction of the results. We identified the type of data source as the main factor that contributes to the heterogeneity, though the type of data source did not explain all the heterogeneity. We suggested that some of the heterogeneity could be attributed to different methods used to ascertain drug discontinuation.  4.2 Study Goals  We theorized that the heterogeneity in reported comparative persistence of TNFα antagonists that we identified was a consequence of the diversity in methods used to ascertain drug discontinuation. To test this hypothesis, we applied a commonly used methodical approach to estimate persistence [17,208]. We ascertained drug discontinuation based on an algorithm that included either switching or a drug-free interval (also known as permissible gap) between dispensing events. The study hypothesis is that increasing lengths of the drug-free interval are associated with a significant change in comparative persistence, which manifests as differences in the significance and/or the direction of these estimates. We also explored patterns of reinitiation of the index drug after a drug-free interval of at least 30 days.     88 4.3 Patients and Methods 4.3.1 Data Source and Study Cohort  We analyzed a cohort of British Columbia residents who received a first course of a TNFα antagonist between March 2001 and December 2008, and had also been diagnosed with RA. Follow-up data were available until December 31, 2009. Patients were identified using four British Columbia Ministry of Health administrative databases: PharmaNet (prescription dispensing data), Medical Service Plan (MSP) registration information (demographic data), MSP Payment Information (fee-for-service payments to physicians and alternative providers), and the Discharge Abstract Database (hospital separations). The databases were linked using a study-unique anonymized identifier. The study protocol was approved by the Clinical Research Ethics Board of the University of British Columbia.  The study cohort was identified based on the algorithm presented in Appendix B, Figure 23, page 270. Patients were included based on one or more recorded dispensing claims for infliximab, adalimumab or etanercept between March 2001 and December 31, 2008. The index date was defined as the date of the first dispensing event of a TNFα antagonist after a run-in period of at least 36 months of follow up without TNFα antagonist dispensing.. RA patients were selected using similar, though not identical, criteria to those used in previous studies of RA patients in British Columbia [101-103]. RA patients were identified based on two outpatient visits with diagnosis of RA at least 60 days apart, or one hospitalization with a recorded discharge diagnosis of RA in three years prior to the index date. A list of International    89 Classification of Diseases (ICD) codes that were used in the study is presented in Appendix B, Table 28, page 271. Diagnostic codes were recorded in the databases for billing purposes and were based on clinical judgment of the care-providing physician. Additionally, patients were required to have continuous provincial MSP coverage three years before the index date. A gap shorter than 30 days was not considered an interruption. Patients were excluded if they were previously treated with anakinra, rituximab or abatacept, if sex or date of birth were missing, if they had a concurrent diagnosis of Crohn’s disease (based on at least one outpatient or inpatient diagnosis code in the three years preceding the index date), or if they were younger than 18 years old at the index date (to remove patients with juvenile RA). 4.3.2 Exposure and Covariates The initial TNFα antagonist (‘the index drug’) was identified for each patient based on their pharmacy dispensing records. We limited the analysis to a single exposure categorical variable for each patient.  The choice of a specific TNFα antagonist drug could be associated with disease severity or additional factors that influence drug persistence. Therefore, we also included demographic and clinical status variables in the model to adjust for possible confounding bias. The covariate variables are presented in Chapter 3, Table 4, page 61. Demographic variables added to the model included sex, age and a proxy variable for socio-economic status (the annual deductible for prescription cost, which was based on family annual income [141]). Clinical status variables included the number of inpatient and outpatient encounters in the year prior to    90 initiating treatment, the duration of disease (defined as time from the first recording of diagnostic code of RA in the data to the index) and the presence and severity of comorbidities. We used Quan’s ICD-9-CM and ICD-10 coding algorithm for administrative databases, excluding rheumatic diseases [175] to determine the Charlson comorbidity score [176] for each patient in the cohort during the year preceding the index date. We also included three variables for other RA drugs: concomitant methotrexate (MTX) based on at least one dispensing in the 200 days prior to index30 (1=yes, 0=no) and dispensing claims of any nonsteroidal anti- inflammatory drug (NSAID) in the year preceding the index date (1=yes, 0=no). The third variable, the number of previous synthetic antirheumatic drugs and corticosteroids was based on the dispensing of the following nine drugs within three years prior to the index date: MTX, hydroxychloroquine, sulfasalazine, azathioprine, cyclosporine, minocycline, penicillamine, sodium aurothiomalate, prednisone and intra-articular triamcinolone/methylprednisolone acetate. We included a categorical variable for the calendar year of the index date, which allowed adjustment for possible secular trends in clinical practice [157-160] and late availability of adalimumab. Finally, we included several Bernoulli variables for the 12 highest volume prescribers (physicians who initiated > 70 courses in the study cohort), to adjust for prescriber propensity for treatment discontinuation. The rationale for considering the covariates is presented in Chapter 2, Section 2.2.3.3, page 41.                                                   30  The period was selected based on mean plus two standard deviations of between-dispensing intervals of MTX in the study cohort.    91 4.3.3 Outcomes The two outcomes variables were (a) a continuous drug persistence variable and (b) a binary censor variable. Drug persistence in years was calculated as the difference between the date of drug discontinuation and the index date. Drug discontinuation was ascertained by either switching to another antirheumatic monoclonal antibody or immunosuppressant drug (infliximab, adalimumab, etanercept, anakinra, rituximab, abatacept, certolizumab and golimumab), or elapsing of a drug-free interval. A drug-free interval was defined as a period without additional dispensing event of the same pharmaceutical component after the days- supply of the latest dispensing was exhausted (Chapter 2, Figure 5, page 22). The date of drug discontinuation was set to the end of days-supply or the date of the first dispensing of a different antirheumatic monoclonal antibody or immunosuppressant drug, whichever was earliest. Patients were considered censored if they were continuously treated with the first TNFα antagonist on December 31, 2009 (end of follow-up period) or when their date of drug discontinuation was during an interruption of more than six days in the provincial MSP coverage. Death and emigration from the province are the most common causes of interruptions in the provincial MSP coverage. For details on assigning the values of the outcome variables refer to Appendix B, Figure 24, page 271. Original data on the number of days-supply recorded in PharmaNet were unreliable (Chapter 2, Section 2.2.4.1, page 47); hence, we also calculated the expected number of days-supply based on estimated vials dispensed, which was imputed using the cost field in the dispensing     92 record. We used the longest duration of days-supply, recorded or calculated, to determine both the length of drug-free interval and the date of drug discontinuation. For discussion on calculating the number of days-supply refer to Chapter 2, Section 2.2.4.1, page 47.  4.3.4 Sensitivity Analysis  In multiple analyses, four lengths of drug-free interval were used to ascertain drug discontinuation: 30, 60, 90 or 180 days after the days-supply was exhausted. Lengths of 30-90 days were commonly used to ascertain discontinuation in other studies of comparative persistence of TNFα antagonists in administrative health databases (30 days [61-63,213,214], 60 days [72,107,213,215,216], 90 days [64,213]). We also applied a length of 180 days, which we considered appropriate to identify patients who discontinued drug instead of only experiencing a temporary interruption in drug persistence. A temporary interruption in drug persistence should be interpreted as poor compliance (or nonpersistence) rather than discontinuation [77]. This length of drug-free interval (180 days) has also been used in estimations of comparative persistence using clinical data [113,114,170]. 4.3.5 Statistical Methods  Summary statistics of baseline characteristics were compiled across the three drug groups. We assumed normality of the continuous variables compatible with large sample size (>2500) and the central limit theorem. The significance of differences between the drug groups was    93 assessed with one-way analysis of variance (ANOVA) F-test for continuous variables and the Pearson’s Chi-square test for categorical variables. All statistical tests were two-sided.  The product-limit method and the log rank test were used to estimate and compare levels of drug persistence. The results were presented for increasing lengths of drug-free interval. Pairwise comparisons between the three drug groups were planned a priori but were considered only when the overall comparison indicated a significant difference (a two-tailed p- value < 0.05) between TNFα antagonists. The crude and adjusted hazard ratios for drug discontinuation were estimated using Cox proportional hazards regression. Three pairwise comparisons were presented: infliximab versus adalimumab, infliximab versus etanercept and adalimumab versus etanercept. Drug-sex and drug-age interactions were examined but were included in the final model only if the overall significance was < 0.1. The proportional hazards assumption was assessed by testing for the significance of interactions between the variables and time or the natural logarithm of time, as well as by plotting the scaled Schoenfeld residuals [177]. The validity of the linearity assumption of continuous variables was assessed by log likelihood tests in models that included either categorical variable or polynomial variables. If nonlinearity was detected, the variable was categorized. All calculations were performed using the SAS software package [178].  4.3.6 Analysis of Reinitiation  We explored patterns of reinitiating the index drug in discontinuers. In these patients, drug discontinuation was ascertained based on an algorithm that included a drug-free interval of 30    94 days or switching. Reinitiation was ascertained when discontinuers reinitiated the index drug (before switching to another antirheumatic monoclonal antibody or immunosuppressant drug). In discontinuers who were followed for at least 60, 90 or 180 days from the date of drug discontinuation, we then estimated proportions of patients who reinitiated the index drug within 31-60 days, 31-90 days or 31-180 days, respectively. Chi-square tests were used to detect statistically significant differences. ‘Added persistence’ was defined as the time from the date of discontinuation based on a drug-free interval length of 30 days or switching until the date of discontinuation based algorithms that included intervals of 60, 90 or 180 days or switching (Figure 11). The product-limit median quantity of ‘added persistence’ in patients who reinitiated treatment was estimated for each drug group.  We considered that the results we obtained in analysis of reinitiation could have been biased in patients treated with adalimumab because of shorter duration of follow-up attributable to being the last of the three TNFα antagonists introduced to the Canadian market (about 3.5 years after infliximab and etanercept). We therefore repeated the analyses of reinitiation in discontinuers, limiting the analysis to patients with index date between January 2005 and December 2008. At that time, all three drugs were available.     95 Figure 11: Calculating "added persistence"          Index date Date of discontinuation based on a drug-free interval of 30 days Drug reinitiation Date of discontinuation based on a drug-free interval of 60/90/180 days  Persistence based on a drug-free interval of 30 days Duration of interruption  ‘Added persistence’    96 4.4 Results 4.4.1 The Study Cohort and Baseline Characteristics Between March 2001 and December 2009, a total of 7,212 British Columbia residents were treated with TNFα antagonists. The study cohort included 2,923 RA patients who initiated TNFα antagonists by December 31, 2008 and met all selection criteria (Chapter 3, Table 5, page 65); 2,104 (72%) were females and age ranged from 18-92 years (mean±SD, 55.9±13.6 years). Patients treated with adalimumab were older and had more previous hospital admissions and lower frequency of concomitant MTX and dispensing claims of NSAIDs compared to patients on infliximab and etanercept. Baseline characteristics are presented in Table 5, page 65. 4.4.2 Effect of Increased Length of Drug-Free Interval The product-limit persistence estimates for different discontinuation ascertainment algorithms are presented in Figure 12 and Table 9. As expected, a longer drug-free interval was associated with improved persistence within each TNFα antagonist and an increase in the overall median persistence from 1.13 years to 2.19, 2.87 and 3.70 years, when applying drug-free intervals of 30, 60, 90, and 180 days, respectively. This suggests that some patients had short interruptions in drug persistence after which they reinitiated the index drug. Applying a drug-free interval of 180 days, persistence on the three individual TNFα antagonists was similar. When applying    97 lengths of 30, 60 or 90 days, infliximab was associated with statistically significant improvements in persistence compared to the other two TNFα antagonists. The persistence curve of patients on adalimumab and etanercept became closer to those on infliximab with increasing of drug-free interval lengths (Figure 12), which suggests that interruptions in drug persistence were more common with adalimumab and etanercept, after which patients reinitiated the index drug. Similarly, the results of univariate Cox regression demonstrated that the hazard ratios for discontinuation approached 1.0 (Table 10, crude hazard ratios) for longer drug-free intervals. For drug-free intervals of 30, 60 or 90 days, infliximab was associated with a statistically significant lower hazard for drug discontinuation compared to the etanercept and adalimumab, with hazard ratios of 0.66, 0.78 and 0.84 compared to etanercept and 0.71, 0.76 and 0.80 compared to adalimumab (for drug-free intervals of 30, 60 or 90 days, respectively). As with the Kaplan-Meier analysis, applying a drug-free interval of 180 days, the hazards for drug discontinuation with infliximab, adalimumab and etanercept were similar.     98 Figure 12: Persistence by algorithm to ascertain discontinuation A. Switching or a drug-free interval of 30 days      99 B. Switching or a drug-free interval of 60 days     100 C. Switching or a drug-free interval 90 days      101 D. Switching or a drug-free interval of 180 days      102 Table 9: Kaplan Meier estimates of persistence for increasing drug-free intervals Drug % censored Median persistence (95% CI) IQR 1 year 5 years A drug-free interval of 30 days All three 25 1.13 (1.02-1.23) 0.32-3.78 52% 20% Infliximab  34 1.76 (1.39-2.12) 0.42-7.61 61% 33% Adalimumab  31 1.00 (0.86-1.23) 0.33-3.90 50% n/a Etanercept  20 1.01 (0.94-1.12) 0.31-3.01 50% 15% Log rank test p-value<0.0001 (infliximab versus etanercept p-value<0.0001, infliximab versus adalimumab p-value=0.0001, adalimumab versus etanercept NSS) A drug-free interval of 60 days All three 38 2.19 (1.99-2.40) 0.56-n/a 64% 34% Infliximab  41 2.87 (2.23-3.78) 0.78-n/a 71% 42% Adalimumab  44 1.98 (1.52-2.80) 0.48-n/a 61% n/a Etanercept  36 2.07 (1.88-2.37) 0.52-n/a 63% 31% Log rank test p-value<0.0001 (infliximab versus etanercept p-value<0.0001, infliximab versus adalimumab p-value=0.002, adalimumab versus etanercept NSS) A drug-free interval of 90 days All three 43 2.87 (2.57-3.17) 0.64-n/a 68% 39% Infliximab  44 3.39 (2.62-4.47) 0.92-n/a 74% 43% Adalimumab  49 2.78 (1.99-3.35) 0.51-n/a 64% n/a Etanercept  42 2.78 (2.52-3.11) 0.60-n/a 67% 37% Log rank test p-value=0.009 (infliximab versus etanercept p-value=0.006, infliximab versus adalimumab p-value=0.01, adalimumab versus etanercept NSS) A drug-free interval of 180 days All three 48 3.70 (3.32-4.16) 0.76-n/a 71% 44% Infliximab  47 3.68 (2.88-4.95) 1.06-n/a 76% 45% Adalimumab  54 3.33 (2.63-4.10) 0.63-n/a 68% n/a Etanercept  47 3.78 (3.31-4.32) 0.70-n/a 70% 43% Log rank test p-value NSS Infliximab (n=620), adalimumab (n=474), etanercept (n=1829)  % - percentage; 1-year - Percentage still on drug one year after the index date; 5-year - Percentage still on drug five years after the index date; CI - confidence interval; IQR - intra-quartile range, presented are the point estimates for the 25 and 75 percentiles of persistence; n/a - cannot be calculated from available data; NSS - not statistically significant (p-value0.05)    103 Adjusting for patient characteristics and other covariates had a significant effect on the estimated hazard ratios when applying a drug-free interval of 90 days. The hazard ratios for infliximab compared to the other TNFα antagonists became insignificant (Table 10). Otherwise, the addition of the other covariates was often associated with hazard ratios that approached the null without a change in the significance. A significant drug-sex interaction was present when comparing infliximab versus etanercept and applying a drug-free interval of 30 days, but not otherwise. No drug-age interaction was significant.   104 Table 10: Hazard ratios for increasing drug-free intervals Comparison Crude hazard ratios (95% CI)  Adjusted hazard ratios (95% CI) A drug-free interval of 30 days     Infliximab versus etanercept  0.66 (0.59-0.74) p<0.0001 Males  0.51 (0.41-0.64) Females 0.74 (0.65-0.85) p<0.0001 p<0.0001 Infliximab versus adalimumab 0.71 (0.62-0.83) p<0.0001 0.70 (0.60-0.83) p<0.0001 Adalimumab versus etanercept 0.93 (0.82-1.05) p=0.21 0.96 (0.84-1.10) p=0.56 A drug-free interval of 60 days Infliximab versus etanercept 0.78 (0.69-0.87) p<0.0001 0.82 (0.72-0.93) p=0.003 Infliximab versus adalimumab 0.76 (0.65-0.90) p=0.001 0.82 (0.69-0.98) p=0.03 Adalimumab versus etanercept 1.02 (0.89-1.16) p=0.83 0.99 (0.86-1.15) p=0.92 A drug-free interval of 90 days Infliximab versus etanercept 0.84 (0.75-0.95) p=0.006 0.89 (0.78-1.02) p=0.09 Infliximab versus adalimumab 0.80 (0.68-0.94) p=0.008 0.86 (0.72-1.04) p=0.12 Adalimumab versus etanercept 1.06 (0.92-1.22) p=0.45 1.03 (0.88-1.20) p=0.70 A drug-free interval of 180 days Infliximab versus etanercept 0.92 (0.81-1.04) p=0.19 0.98 (0.85-1.13) p=0.76 Infliximab versus adalimumab 0.87 (0.73-1.03) p=0.42 0.95 (0.78-1.15) p=0.57 Adalimumab versus etanercept 1.06 (0.92-1.23) p=0.10 1.04 (0.88-1.22) p=0.68 CI- confidence interval; p – p-value (Cox regression)    105 4.4.3 Analysis of Reinitiation  There were 2,167 discontinuers who were followed for at least 60 days, including 2,153 for at least 90 days and 2,097 for at least 180 days from the date of discontinuation (Table 11). A larger proportion of patients reinitiated etanercept within 60, 90 and 180 days after the beginning of an interruption of 30 days or longer in drug persistence compared to infliximab and adalimumab. This may suggest better compliance with infliximab, manifested as less short interruptions in drug treatment, compared to etanercept. The quantity of ‘added persistence’ was similar across the three drug groups, but it was proportionally smaller with infliximab, because infliximab was associated with longer persistence when applying the algorithm with a 30-day drug-free interval. No differences were observed in drug reinitiation patterns between infliximab and adalimumab. Next, we established a sub-cohort of patients with index dates between January 2005 and December 2008. We first compared drug persistence using nonparametric survival analysis and found the comparative persistence sensitive to the length of drug-free interval (Appendix C, Figure 25, page 276). A significant difference in persistence between drugs was demonstrated when applying drug-free intervals of 30 and 60-day (log rank test p-value <0.0001 and 0.03 respectively) but not when applying 90 or 180-day lengths. This supports the robustness of the results presented in Section 4.4.2, page 96. We repeated the analyses to explore patterns of reinitiation. A significantly larger portion of adalimumab and etanercept discontinuers reinitiated the index drug compared to infliximab (Table 12). Again, no significant difference was observed with the quantity of ‘added persistence’, except when applying the 180 days interval in ascertainment of discontinuation.    106 Table 11: Analysis of patients reinitiating the same drug (treatment initiated 2001-2008) Drug N Number of patients who reinitiated the index drug (%) ‘Added persistence’  median (95% CI) Returned within 31-60 days after discontinuation Infliximab 404 108 (26.7) 2.91 (1.70-4.67) Adalimumab 319 101 (31.7) 1.83 (1.52-3.01) Etanercept 1446 605 (41.8) 2.35 (1.93-2.88) Testing for homogeneity of the three drug groups  (Chi-square) three drugs p<0.0001;      Infliximab vs. etanercept p<0.0001;    Infliximab vs. adalimumab NSS;   Adalimumab vs. etanercept p=0.0008 (log rank test) NSS Returned within 31-90 days after discontinuation  Infliximab 400 137 (34.3) 3.23 (2.19-4.67) Adalimumab 315 119 (37.8) 2.64 (2.13-n/a) Etanercept 1437 709 (49.3) 3.39 (2.79-4.01) Testing for homogeneity of the three drug groups  3 drugs p<0.0001;    Infliximab vs. etanercept p<0.0001;    Infliximab vs. adalimumab NSS;   Adalimumab vs. etanercept p=0.0002 NSS Returned within 31-180 days after discontinuation  Infliximab 390 146 (37.4) 1.30 (2.78-n/a) Adalimumab 302 131 (43.4) 3.44 (2.34-n/a) Etanercept 1405 771 (54.9) 5.23 (4.30-6.40) Testing for homogeneity of the three drug groups  3 drugs p<0.0001;    Infliximab vs. etanercept p<0.0001;    Infliximab vs. adalimumab NSS;   Adalimumab vs. etanercept p<0.0003 NSS N - number of patients who discontinued the drug based on a drug-free interval of 30 days or switching; n/a – cannot be estimated from available data; NSS – not statistically significant (p-value ≤0.05); p – p-value; vs.- versus     107 Table 12: Analysis of patients reinitiating the same drug (treatment initiated 2005-2008) Drug N Number of patients who reinitiated the index drug (%) ‘Added persistence’  median (95% CI) Returned within 31-60 days after discontinuation Infliximab 125 25 (20.0) 1.79 (1.06-n/a) Adalimumab 286 92 (32.2) 2.14 (1.72-n/a) Etanercept 656 241 (36.7) 2.30 (1.74-n/a) Testing for homogeneity of the three drug groups  (Chi-square) three drugs p=0.001;    Infliximab vs. etanercept p=0.0003;    Infliximab vs. adalimumab p=0.01;   Adalimumab vs. etanercept NSS (log rank test) NSS Returned within 31-90 days after discontinuation  Infliximab 121 34 (28.1) 1.79 (1.10-3.24) Adalimumab 282 108 (38.3) 3.44 (2.14-n/a) Etanercept 650 291 (44.8) 3.55 (2.65-n/a) Testing for homogeneity of the three drug groups  3 drugs p=0.002;    Infliximab vs. etanercept p=0.0005;    Infliximab vs. adalimumab p=0.05;   Adalimumab vs. etanercept NSS  NSS Returned within 31-180 days after discontinuation  Infliximab 115 34 (29.6) 1.99 (1.18-n/a) Adalimumab 269 117 (43.5) 3.44 (2.16-n/a) Etanercept 630 314 (49.9) n/a (n/a) Testing for homogeneity of the three drug groups  3 drugs p=0.0002;    Infliximab vs. etanercept p<0.0001;    Infliximab vs. adalimumab p=0.01;   Adalimumab vs. etanercept NSS  0.03 %- percentage; N - number of patients who discontinued the drug based on a drug-free interval of 30 days or switching; n/a - cannot be estimated from available data; NSS - not statistically significant (p-value ≤0.05); p - p-value; vs.- versus     108 4.4.4 Effect of Covariates  Sex, age, the number of previous hospital admissions, prescriber propensity for treatment discontinuation and concomitant of MTX were all significantly associated with TNFα antagonist persistence. Females, the extremes of age range (<30 or >70 years old) and increased number of admissions were associated with increased risk of drug discontinuation. Concomitant MTX was associated with improved persistence. Comorbidities and disease duration had no effect on drug persistence in analyses of all four drug-free interval lengths. However, the significance of the association of the remaining variables with persistence (the calendar year of index date, the number of visits, the annual deductible level, dispensing claims for NSAIDs and the number of previous synthetic antirheumatic drugs and corticosteroids) was sensitive to the drug-free interval length. 4.5 Discussion Our results support the study hypothesis that the heterogeneity in reported comparative persistence of TNFα antagonists was caused by the use of different algorithms to ascertain drug discontinuation.  In this study, we evaluated the effect of a drug-free interval, which is a common end-point in algorithms to ascertain drug discontinuation in analyses of administrative data. The length of drug-free interval used in an algorithm, which is an exogenous factor chosen by the investigator, was an important factor affecting both the magnitude and the statistical significance of the hazard ratios for discontinuation of TNFα antagonists in patients with RA.    109 4.5.1 Ascertainment of Drug Discontinuation  Ascertainment of discontinuation is critical in studies measuring drug persistence, because persistence is measured as time to discontinuation of a given drug. Yet investigators using clinical data have typically not reported the method used in ascertainment of drug discontinuation in studies of persistence with TNFα antagonists in RA patients. Instead, ambiguous descriptions have been provided, such as "The date of drug discontinuation was recorded" , “Withdrawal from treatment was registered prospectively” [150], “All treatment terminations ...are recorded” [115]. Comparison between studies is not feasible without full disclosure of how drug discontinuation was ascertained and the date of discontinuation was defined. In analyses of dispensing claim data, drug discontinuation was often ascertained based on two events: (a) elapsing of a drug-free interval (‘permissible gap’) after the exhaustion of the drug dispensed [17] or (b) switching to an alternative drug (either from the same therapeutic class or from a different class). The rationale for applying a drug-free interval in ascertainment of drug discontinuation, rather than the exhaustion of the dispensed days-supply, was to minimize underestimation of persistence because patients treated in clinical setting were expected to have variable times to refill after the dispensed quantity was exhausted [217,218]. This could have been caused by dosing or recording errors, dose adjustment, stockpiling, mild adverse events such as abnormal laboratory results, drug interactions, vacations, hospitalizations and so on. Ascertainment of discontinuation using a drug-free interval (sometimes also switching) allows investigators to assign persistence a continuous or Bernoulli value (Chapter 2, Section    110 0, page 15). Assigning continuous values to persistence permits survival analysis, which accounts for censored patients and reflects lengths of follow up for all patients.  Selecting the length of the drug-free interval presents a major challenge and as a result, different lengths were applied in different studies of the same treatment. In published analyses of administrative data, various lengths of drug-free interval were used to ascertain discontinuation of TNFα antagonists in RA patients: 30 days [61-63,213,214], 60 days [72,107,155,213,215,216] and 90 days [64,213]. Three principles should be considered in the decision on the lengths of drug-free interval. First, the selection of the length of the drug-free interval should be based on the pharmacological properties of the drug and the clinical situation [7,59]. The rationale for selecting a specific interval was rarely discussed in studies of TNFα antagonists or other drugs, whereas the interval was often selected to match previous studies of the same drug. For example, a drug-free interval of 30 days was commonly used for daily drug [18,219] based on convention rather than pharmacological rationale. Ideally, the length of a drug-free interval should approximate the maximum period for which deprivation of the drug would not reduce therapeutic benefit (or increase harm) [220]. Unfortunately, this knowledge is limited for most treatments. In patients treated with infliximab at a dose of one milligram per kilogram body weight with concomitant MTX, more than 60% of the patients maintained the 20% Paulus    111 response criteria31 for a median duration of 16.5 weeks and the 50% Paulus response criteria for 12.6 weeks [75]. With adalimumab, response (based on the European League Against Rheumatism (EULAR) 32 and American College of Rheumatology (ACR)33) was sustained four weeks to three months [76]. However, large variation was observed between patients and the applicability of the response criteria used to clinical outcome (such as pain, morbidity and mortality) is not established. Using a short drug-free interval (30 days or less) in studies of TNFα antagonists has therefore led to biased and underestimated duration of therapeutic effect.  Second, sensitivity analysis applied to various lengths of drug-free interval is recommended for checking the robustness of persistence estimates [30,59,73]. Often the selection of the length of drug-free interval to ascertain discontinuation was arbitrary because of limited knowledge. It is intuitive that a longer drug-free interval is associated with improved persistence on an individual drug, because some patients reinitiate the index drug after a short interruption and                                                  31  Paulus criteria were developed based on data from multiple trials in RA patients, according to their ability to differentiate between patients treated with DMARDs and patients treated with placebo. Response requires fulfillment of at least four of the following criteria: 20% or 50% (a) improvement in morning stiffness, (b) erythrocyte sedimentation rate (ESR), (c) joint tenderness score, (d) joint swelling score and (e) improvement by at least two grades on a five-grade scale (or from grade two to grade one) for patient and physician global assessments of current disease severity [481]. 32  EULAR response specifies response based on disease activity score in 28-joints (DAS28). Good responders are patients with an improvement in DAS28 of >1.2 and a present score of ≤3.2. Moderate responders are patients with an improvement of >0.6–≤1.2 and a present score of ≤5.1, or an improvement of >1.2 and a present score of >3.2 and non-responders are patients with an improvement of ≤0.6, or patients with an improvement between 0.6–1.2, and a present score of >5.1. DAS28-defined remission is a score of <2.6 [439]. 33  The ACR response criteria are described in Appendix A, Section A.3.1, page 3.    112 this interruption is not considered to be a discontinuation. A longer drug-free interval was associated with improved persistence in our study and in studies of other drug [218,221-223]. When considering comparative persistence, the effect of increasing lengths of drug-free interval is not straightforward. If the effect is similar for both drugs, a change in the drug-free interval applied to ascertain discontinuation does not affect the estimates of comparative persistence. If, on the other hand, noncompliance and short interruptions are more frequent with one drug, then increased interval may be associated with differences in magnitude, significance and even direction of comparative persistence. Sensitivity analysis applying varied lengths of drug-free interval is helpful to disclose instability of the comparative estimates. Only a minority of studies estimating comparative persistence with different therapeutic classes performed sensitivity analysis.  Greevy et al 2011 [217], in their analysis of oral antidiabetic drugs, showed that increasing the length of drug-free interval from zero to 30 days reversed the direction of the hazard ratio for discontinuation of sulfonylurea versus metformin. More commonly, sensitivity analysis showed that estimates of comparative persistence were robust, in survival analysis (persistence as a continuous variable) or logistic regression (persistence as a Bernoulli variable) analyses. Increasing lengths of drug-free interval were associated with robust estimates of comparative persistence in patients treated with  cholinesterase inhibitors for Alzheimer’s disease [224], interferons and glatiramer acetate for multiple sclerosis [225] or with serotonin-specific reuptake inhibitor for psychiatric conditions [226]. Additionally, an increased drug-free interval did not affect the results of comparisons of six different therapeutic classes of drug used by patients with a variety of chronic diseases [147].     113 In investigations of RA patients treated with TNFα antagonists, only three studies reported persistence estimates when increasing the length of drug-free interval (Appendix C, Table 29, page 280). The first two studies examined percentage of discontinuers in the first anniversary after the index date, using administrative data. Li et al 2010 [213] reported discontinuation in patients treated with infliximab and etanercept. The investigator reported “comparable discontinuation rates” of the two drugs (an unknown test for between group comparisons was performed and the differences were insignificant for all lengths applied). Ogale et al 2011 [215] examined the proportion of discontinuers based on switching and a drug-free interval of 60 and 180 days. They demonstrated that longer drug-free intervals were associated with decreased percentage of patients who experienced discontinuation, “but the relative pattern remained the same”. No statistical test was reported. The third study examined discontinuation at the second anniversary after the index date. Schmeichel-Mueller et al [216] applied drug- free intervals of 60 and 360 days in ascertainment of discontinuation using administrative data. This study was reported as an abstract and comparisons between discontinuation proportions in the three drug groups were not reported. Overall, the effect of increasing drug-free interval lengths on estimates of comparative persistence remained unclear based on these studies. The third principle in ascertainment of drug discontinuation pertains to the interpretation of the outcome measure. Specifically, the length of the drug-free interval determines the implication of the outcome measure estimated in the study. Prolonging the drug-free interval ensures a    114 lower frequency of reinitiating the same drug after ascertainment of discontinuation; therefore, the measure estimated is persistence. When the investigators apply a short drug-free interval of 60 days or less, the measure estimated encompasses both duration and intensity of dispensing events34. Therefore, the outcome measure estimated in these studies should be regarded as a mixed measure of compliance and persistence. Furthermore, the different methods used to measure the outcome ‘persistence’ in different studies could imply that investigators in fact measured different phenomena and therefore (a) heterogeneity in the results is expected, even between analyses of the same type of data source and (b) pooling the results from different studies is inappropriate. 4.5.2 Possible Superiority of Infliximab We showed that the proportion of patients reinitiating the index drug after a temporary interruption probably played a role in the differential influence of the length of drug-free interval on the estimated persistence with different TNFα antagonists. As a result, the significance of comparative persistence estimates was sensitive to the length of the drug-free interval. We interpret interruptions in drug persistence followed by reinitiating the index drug                                                  34  Caentano 2006 [59] was the first to distinct between duration and intensity in following the prescription recommendations.    115 as noncompliance, rather than nonpersistence35. The reasons for temporary interruptions in drug persistence were previously studied in RA patients who were treated with etanercept [227]. The main reasons identified were neutropenia, infection, abnormal liver function tests and thrombocytopenia. We found that interruptions in drug persistence followed by reinitiating the index drug were less frequent on infliximab. Our results are supported by an abstract presented by Schmeichal- Mueller 2011 [216]. In an analysis of administrative data, a higher proportion of etanercept (62.1%) and adalimumab users (49.7%) reinitiated the index drug after an interruption of longer than 60 days compared to infliximab (12.8%). Flendrie et al 2009 [228] showed that temporary interruptions in persistence were more common with adalimumab (42.6%) compared to infliximab (19.3%) or etanercept (13.3% of the treated patients). The most common reason was adverse events (not further specified). Unfortunately, temporary interruption was not defined in the report. Better compliance with infliximab compared to etanercept was also reported by Harley 2003 [229] and Li 2010 [213], and compared to etanercept and adalimumab by Carter 2010 [230] and Schmeichel-Mueller 2011 [71]. There are several possible explanations for possible improved compliance with infliximab. First, adverse events were shown to predominate in infliximab discontinuation, compared to other TNFα antagonist drugs. In RCTs, withdrawals due to adverse events (WDAEs) were                                                  35  For discussion on the definition of persistence and compliance, refer to Chapter 2, Section 2.1.1, page 3.    116 more common with infliximab compared to adalimumab [231]. Other studies showed that in routine clinical practice, a higher proportion of infliximab discontinuations were due to adverse events36, especially acute systemic reactions [108,110,170]. It is unlikely therefore, that patients who discontinued a drug because of an adverse event, especially a serious one, reinitiated the same drug. Second, infliximab is an infusible agent and regular physician follow-up is required for administration. Regular encounters have been shown to encourage compliance and persistence of drug therapy [162,184,185], though this was not studied in patients treated with TNFα antagonists. Lastly, dose escalation is more frequent in patients treated with infliximab compared to the other drugs [62,72]. It is possible that the use of standard dosing to determine the number of days-supply led to an overestimation of this number and an underestimation of the length of actual drug-free intervals in patients who were treated with escalating doses of infliximab (Chapter 2, Section 2.2.4.2, page 52).  Our results suggest possible superior adherence with infliximab. Improved persistence with infliximab was observed also by Gomez-Reino et al [117], Tang et al [104], Yazici et al [62] and recently by Ogale et al 2011 [215]. This may be the result of more frequent dose                                                   36  In meta-analysis of multiple studies examining reasons for discontinuing the TNFα antagonists in RA, similar proportion of patients discontinue infliximab due to adverse events, compared to adalimumab and etanercept (Appendix A, Table 24, page 3); nonetheless, the included studies were not identified using systematic review methodology.     117 escalation in patients on infliximab as an alternative to switching in patients with lack or loss of efficacy [62,72] (Chapter 2, Section 2.2.4.2, page 52), or, as mentioned before, due to frequent physician encounters.   4.5.3 Persistence as a Measure That Integrates Benfit and Harm The assumption that drug persistence could be used as an indirect measure of therapeutic benefit-herm balance, first introduced by Wolfe in 1995 [45], has often been quoted in studies of rheumatologic diseases, especially RA [64,232-234]. Correspondingly, comparative persistence, a relative measure, was suggested as an indirect measure of the relative benefit and harm of individual drugs.  An advantage of using drug persistence is that benefit and harm are inherently accounted for together because the main reasons for discontinuations are reduced therapeutic benefit or perceived harm. On the other hand, benefit and harm are usually reported separately in clinical trials. This is important because clinical decisions have inherent tradeoffs between positive and negative drug effects; therefore, a measure that reflects both benefit and harm is easier to interpret and implement in clinical practice compared to separate measures of either benefit or harm [235]. In addition, persistence can be estimated easily and cheaply, using varying types of data source, including administrative data.     118 However, we advise caution when interpreting drug persistence as a measure of therapeutic benfit-harm balance. We mentioned before (Chapter 3, Section 3.5.1, page 72) the absence of direct evidence as to the validity of persistence as a measure of benefit and harm. Longitudinal studies estimating an association between a change in disease severity or an adverse event during treatment and subsequent hazard for drug discontinuation are needed. An alternative approach, relating therapeutic outcomes to persistence at a single time point could be biased due to the analytic methods used. Intention to treat analysis (ITT) with last observation carried forward (commonly used in RCTs) or per protocol analysis [156,236]) tend to inflate the proportion of patients who experienced therapeutic benefit. On the other hand, ITT analysis per se [156,237] tends  to underestimate the therapeutic benefit, as observations later than the discontinuation date are included. Drug discontinuation represents cessation of the drug opportunity to impact therapeutic effect, and therefore patients who discontinue the drug are expected to experience worsening of disease activity, especially if the drug is effective.   A second limitation in interpreting persistence is that it is a summary measure, and as such it failed to express the diverse reasons that cause discontinuation and weight these causes accordingly. Discontinuations caused by insufficient therapeutic benefit are accounted for in a similar way as is discontinuation due to harm regardless of severity. Drug persistence was suggested to indirectly estimate benfit-harm balance based on the assumption that drug is continued in these patients provided that it has therapeutic benefit and only mild harmful effects that the benefit overweights [45,209]. Theoretically, three conditions required to be fulfilled.     119 1. All patients who discontinue the drug did so as a result of either inadequate therapeutic benefit (real or perceived inefficacy) or drug-related harm (real or perceived). Based on previous studies, decreased benefit and perceived harm are the main reasons for discontinuing TNFα antagonists in RA (Appendix A, Section 0, page 255), however, other causes should be considered as well. A notable proportion of the patients (up to 23%) discontinued the drug due to other reasons, such as planned pregnancy [108,170,180] and financial constraint [68]. Drug persistence may be affected by nonmedical reasons including out-of-pocket costs, change in insurance coverage or formulary status, preference or aversion to different routes of drug administration and advertising or new products choice [235]. Most importantly, drug persistence does not account for discontinuation due to remission in disease activity [170,179]. Some guidelines suggest that in RA patients who achieve remission with TNFα antagonists, discontinuation of these drugs could be considered [238,239]; even though a high relapse rate is expected in patients who discontinue TNFα antagonists due to remission [240-242]. Discontinuation followed by relapse in disease activity and reinitiation of a drug leads to a temporary interruption in drug persistence. Comparative persistence, as a relative measure, could theoretically reflect relative benefit- harm balance accurately if discontinuation due to non-effectiveness-related reasons is distributed equally in the drug groups. One of the implications of using a relative measure, is that systematic errors in measuring the outcome (in this case non-effectiveness-related reasons for discontinuation) cancel each other out as long as they are equally distributed between the groups. Imbalances between drug groups in these types of nonmedical causes of discontinuation however, may bias the results. For example, of the three TNFα    120 antagonist drugs we studied, patient preference for a subcutaneous route of administration [181] would only be a reason for infliximab discontinuation and as a causal factor may have exerted a greater effect at the time when a new subcutaneous drug (adalimumab) was introduced.  As a result, infliximab persistence may have fluctuated during the study period, and comparative persistence may therefore have underestimated the effectiveness of infliximab compared to the other two drugs. 2. All patients who experienced harm or inadequate benefit discontinue the drug. This is not true in treatment with TNFα antagonists. Not all RA Patients who experience inefficiency or harm discontinue the TNFα antagonists. Kievit et al 2009 [243], for example, showed that only 19% of the patients who were classified as nonresponders after three months of treatment with TNFα antagonists discontinued these drugs, in concordance with the Dutch guidelines. Alternative strategies to inadequate therapeutic benefit include increasing dose (especially of infliximab) [244], or adding-on or optimization of concomitant synthetic antirheumatic drugs [245,246]. In case of harm, the drug may be temporarily interrupted [227] or the dose reduced [247,248].  Comparative persistence could reflect relative benefit-harm balance accurately if other strategies for lack of benefit or perceived harm are equally distributed between the drug groups. However, this is not the case with RA patients treated with TNFα antagonists. Previous studies demonstrated higher prevalence of dose adjustment in RA patients treated with infliximab compared to the other two TNFα antagonists (Appendix A, Section A.3.6.2, page 265). These patients persist with the TNFα antagonists despite ineffectiveness. Some patients may experience benefit (or decreased drug-related harm)    121 with the new dose schedule, but in other patients, these strategies may only delay discontinuation until the effect of the new dose schedule is clear. Consequently, persistence might overestimate effectiveness in patients treated with infliximab and comparative persistence of infliximab versus other drugs could be biased. 3. Only drug properties and patients characteristics determine persistence. We present evidence that this condition is not fulfilled. Prescriber propensity for treatment discontinuation was demonstrated to affect measures of persistence and comparative persistence (Chapter 3, Section 3.4, page 64). This is probably a result of discrepancies in clinical guideline regarding indication for TNFα antagonist discontinuation, especially the use of sometimes vague definitions of lack of efficacy or harmful events (Appendix A, Section A.3.2, page 251). More importantly, neither the Canadian nor the American guidelines defined inefficacy or harm [249-253], and therefore discrepancies in discontinuation decisions are expected amongst British Columbian prescribers who may follow one or other guideline. In Chapter 3, we showed that not only do individual prescribers presented significantly different rates of discontinuation compared to others (Table 7, page 69), but prescriber propensity for treatment discontinuation also confounded the estimates of comparative persistence (Table 8, page 71). Some prescribers had a different baseline tendency to recommend discontinuation compared to most prescribers. Similarly, Zhang et al showed the effect of the prescriber propensity for treatment discontinuation on TNFα antagonists’ discontinuation related to decreased benefit in a recent study [48]. This study is discussed in Chapter 5, Section 5.5.1.3, page 153.     122 In this chapter (Chapter 4), we demonstrate an exogenous factor that influences estimates of persistence and comparative persistence – the investigator. The instability of persistence and comparative persistence estimates and the effect of methods on these estimates also limit their validity as measures of benefit-harm balance. Briefly, persistence and comparative persistence have attracted interest as measures of benefit- harm balance, especially in rheumatic diseases. They are easy to estimate and express the effect of both therapeutic benefit and harm. We discussed multiple limitations of the measures, and showed that the interpretation of either persistence or comparative persistence as a measure of benefit-harm balance may be misleading and caution should be taken. 4.5.4 Strengths and Limitations of the Study Please refer to Chapter 3, Section 3.5.3, page 77.  4.5.5 Clinical and Health Policy Implications We showed that the estimates of comparative persistence with TNFα antagonists in RA were sensitive to the algorithm used to ascertain drug discontinuation. Previous studies were commonly conducted using a short drug free interval of 30 [61-63,213,214] or 60 days [72,107,213,215,216], and sensitivity analysis for the robustness of the etimates was not performed. Therefore, we regard published results of previous studies to be problematic.    123 Distinctions between these TNFα antagonists based on available evidence from persistence studies are not advised. Instead of using comparative persistence with TNFα antagonists as a measure of relative effectiveness estimates  for clinical practice or health policy decision- making, we suggest postponing use until methodological issues are resolved and more reliable estimates are possible.  To improve the quality and applicability of research, we believe a consensus between clinicians, policy-makers and researchers on the algorithm to ascertain discontinuation is necessary. This consensus should also apply to studies measuring harm and other outcomes of TNFα antagonist treatment, since the risk is estimated until drug discontinuation or prespecified time after the discontinuation. We suggest considering the complex pharmacologic properties and the effect of the drugs before reaching a consensus on the most appropriate algorithm. The median duration of effect of infliximab in RA patients, for example, was 16.5 weeks (intra-quartile range 7 to >20 weeks) after administration [75] and the effect of adalimumab was shown to be sustained 4 weeks to 3 months [76]. Both estimates of effect duration exceeded drug exposure period based on the drug-free interval of 30 days, which is commonly used in observational studies.  4.5.6 Research Implications and Directions We have suggested that sensitivity of persistence and comparative persistence estimates to the method used to of ascertain discontinuation has implication for interpretability of other therapeutic outcomes in RA patients treated with TNFα antagonists. Investigators defining    124 TNFα antagonist exposure based on dispensing data often consider the patients to be exposed based on the days-supply plus a drug-free interval. This method was used in studies examining serious infection [254], fatal infections [255], hospitalizations [256], congestive heart failure [257], malignancy [255] and other outcomes [258].  To examine the robustness of therapeutic outcomes, we suggest that sensitivity analysis using increasing lengths of drug-free interval should be applied in future comparisons between the TNFα antagonists and in studies of other treatments.  Generally, in the process of synthesizing and translating available knowledge, critical appraisal of the methodology of the studies is warranted. We have shown that that heterogeneity in comparative persistence estimates may be caused by the different methods used to ascertain drug discontinuation. Ascertainment of discontinuation and discontinuation date are especially important in studies of comparative persistence with the TNFα antagonists, because administration schedules vary from every three days (for etanercept 25 mg) to as much as 56 days (for infliximab, third dose and onwards). Using the last administered dose as the date of discontinuation, for example, biases the results toward inferior persistence on infliximab. We did not study the effect of ascertainment of discontinuation date on the direction, magnitude and significance of hazard ratios, and suggest that investigation of these additional hypotheses would yield useful results as well.  We had no access to clinical data; therefore, we could not explore the effect of methods used to analyze this data type on the result obtained, nor could we compare methodologies used in analysis of administrative data to those used for clinical data. Further research using clinical data (clinical records or disease registry) is recommended to support our finding of the    125 significance of the method used to ascertain drug discontinuation in comparisons of persistence with TNFα antagonists. We speculate that discontinuation and the date of discontinuation were inconsistently ascertained in clinical studies, because they are based on physician recording of discontinuation and inconsistency in physician recording is possible. To obtain more robust estimates of comparative persistence for clinical and health policy decision making, it would be necessary to conduct a sensitivity analysis study using clinical data to more accurately establish the date of discontinuation, using date of last administered dose, date of end of drug coverage and date of physician visit to establish drug discontinuation. In addition, comparison of the clinical records to claim/pharmacy records would be required to provide further support for the hypothesis that distinct methods to establish date of drug discontinuation were used in studies using different types of data, and the diversity in methods led to heterogeneity in estimated of comparative persistence. 4.6 Conclusions  Persistence with TNFα antagonists has shown promise as a valid tool to study relative benfit-harm balance in real life RA patients, because the main reasons to drug discontinuation were decreased benefit and perceived harm. However, persistence and comparative persistence should not be considered valid measures of the benfit-harm balance because of the following reasons:     126  In previous observational studies, in about a quarter of the patients who discontinued TNFα antagonists non-effectiveness-related reasons were quoted, and these reasons were not evenly distributed between patients treated with different drugs   Alternative clinical approaches were used in patients who experience lack of benefit or experience harmful events, and these approaches were not evenly distributed between patients treated with different drugs   We showed that both the prescriber and more important – the investigator determine persistence and comparative persistence, and not only drug properties and patients characteristics    We studied the significance of the length of ‘drug-free interval’ used to ascertain drug discontinuation and demonstrated its impact on hazard ratios for drug discontinuation. It would appear that the heterogeneity in the existing body of evidence on comparative persistence of TNFα in RA is largely a result of lack of standardization of this important design feature of comparative persistent studies. This problem is not easily resolved by one research group therefore a consensus in the field on this critical point would probably be necessary before persistence studies yield the useful findings anticipated. The consensus will allow researchers to conduct research that would contribute to clinical and policy meaningful knowledge to accumulate.      127  Different methods to establish drug discontinuation (sensitivity analysis) should always be used to check the robustness of the results in studies comparing drug persistence. Similar approach (sensitivity analysis) is required in ascertainment of drug exposure in studies of other therapeutic outcomes, to ensure the robustness of the results, because we showed a differential influence of the method to ascertain drug discontinuation on the different drug.  Patients treated with etanercept and adalimumab are more likely than patients treated with infliximab to experience temporary interruption in drug persistence, after which they reinitiate the same drug. This finding indicates better compliance with infliximab, probably due to more frequent encounters with health professionals (compared to patients treated with self-administered drug) and nonself administration of the drug. Consequently, programs to improve compliance with TNFα antagonists should focus on patients treated with etanercept and adalimumab. Additionally, poor compliance with etanercept and adalimumab should be accounted for in budget planning, cost effectiveness studies and so on.    128 CHAPTER 5:  PRESCRIBER PREFERENCE FOR TUMOUR NECROSIS FACTOR ANTAGONISTS PREDICTS PERSISTENCE ON TREATEMNT IN RHEUMATOID ARTHRITIS 5.1 Background Patient characteristics and therapeutic outcomes (both beneficial and harmful) have been established as major determinants of drug persistence in multiple chronic noncurable diseases, including rheumatoid arthritis (RA) [79,232,259,260]. We were interested in persistence with tumour necrosis factor alpha (TNFα) antagonists in patients with RA. A minority of TNFα antagonist persistence studies also assessed non-patient characteristics: prescriber specialty [104], payer type [62,104] and treatment center [110]37. Potential roles of physician preference for the prescribed drug or prescribing habits as determinants of drug persistence have not previously been studied, in RA or other diseases.                                                   37  Persistence studies of other treatments sometime included prescriber characteristics, such as age [135,162], gender [135,162], number of patients treated by this prescriber [135,162] or  specialty [163-166]  in the model.    129 We suspected that TNFα antagonist therapy in RA patients might be especially sensitive to drug preference of the prescribing physician for two main reasons. First, during the study period (2001-2008) there was limited clinical evidence on the comparative effectiveness of the drugs (Appendix A, Section A.3.1, page 241). Second, the published indications for discontinuation of TNFα antagonists were vague and confusing (Appendix A, Section A.3.2, page 251) and care-providing physicians could reasonably be expected to reach different clinical decisions given the same clinical situation (further discussion in this chapter, Section 5.5.1.3, page 153). Consequently, the decisions about which of the TNFα antagonists to initiate first (drug selection) and when to discontinue the drug are likely subject to the prescriber individual preference.   In Chapter 3, we demonstrate the importance of the prescriber propensity for treatment discontinuation both as a predictor of persistence and as a possible confounder of comparative persistence (Section 3.4, page 64). The prescriber recorded on the first dispensing claim for TNFα antagonist was used as a proxy of the care-providing physician. In this chapter, we extend our study of the prescriber’s role in persistence and reveal the effect of prescriber preference for the prescribed drug (PPD) using data on previous prescribing habits. 5.2 Objectives and Hypothesis The main study hypothesis was that increased PPD was associated with improved persistence. We also speculated that the PPD would not confound comparative persistence measures. Therefore, the goals of the study were (a) to estimate the effect of PPD on the TNFα antagonist    130 persistence in patients with RA, (b) to test whether PPD is a possible confounder of estimated of comparative persistence estimates and (c) to characterize and compare levels of PPD in patients who switched to a second TNFα antagonist. 5.3 Patients and Methods 5.3.1 Data Source and Study Cohort  We analyzed a cohort of British Columbia residents who received a first course of a TNFα antagonist between March 2001 and December 2008, and had also been diagnosed with RA. Follow-up data were available until December 31, 2009. Patients were identified using four British Columbia Ministry of Health administrative databases: PharmaNet (prescription dispensing data), Medical Service Plan (MSP) registration information (demographic data), MSP Payment Information (fee-for-service payments to physicians and alternative providers), and the Discharge Abstract Database (hospital separations). The databases were linked using a study-unique anonymized identifier. The study protocol was approved by the Clinical Research Ethics Board of the University of British Columbia.  The study cohort was identified based on the algorithm used for the studies in Chapter 3 and Chapter 4 and presented in details in Appendix B, Figure 23, page 270. Patients were included based on one or more recorded dispensing claims for infliximab, adalimumab or etanercept between March 2001 and December 31, 2008. We used the first dispensing claim    131 for a TNFα antagonist (index dispensing) after a run-in period of at least 36 months of follow up without TNFα antagonist dispensing to identify the index drug, index date and index prescriber for each patient in the cohort. RA patients were selected using similar, though not identical, criteria to those used in previous studies of RA patients in British Columbia [101- 103]. RA patients were identified based on two outpatient visits with a diagnosis of RA at least 60 days apart, or one hospitalization with a recorded discharge diagnosis of RA within three years prior to the index date. A list of International Classification of Diseases (ICD) codes that were used in the study is presented in Appendix B, Table 28, page 271. Diagnostic codes were recorded in the Ministry of Health databases for billing purposes and were based on clinical judgment of the care-providing physician. Additionally, patients were required to have continuous provincial MSP coverage three years before the index date. A gap shorter than 30 days was not considered an interruption. Patients were excluded if they were previously treated with anakinra, rituximab or abatacept, if sex or date of birth were missing, if they had a concurrent diagnosis of Crohn’s disease (based on at least one outpatient or inpatient diagnostic code in the three years preceding the index date), or if they were younger than 18 years old at the index date (to remove patients with juvenile RA). In the current study, we also excluded patients treated by an index prescriber who cumulatively initiated less than five courses in the RA cohort (a ‘low volume’ prescriber).    132 5.3.2 Exposure and Covariates The PPD for a specific patient was determined at the index date and coded as a Bernoulli variable. For each patient, we identified courses of TNFα antagonists initiated by the index prescriber for all patients in the study cohort during the year preceding the index date (excluding the course of interest). From this one-year data, we calculated the proportion of patients with the same index drug divided by the total number of patients prescribed by the same prescriber (Figure 13). A PPD was assigned the value ‘higher’ when the calculated proportion was 60% or higher and the value ‘lower’ otherwise38. A threshold of 60% was selected to ensure that higher PPD meant the prescriber clearly favored this drug, even if only two TNFα antagonists were available (until October 2004) or prescribed. Assigning values based on the prior year allowed us to account for change in preference over time and for the later availability of adalimumab (at the end of 2004, while the other two drugs had been available since 2001). A sensitivity analysis was conducted to determine the robustness of results using thresholds of 70% or 80% for the level of PPD.  Multiple covariates, mainly patient characteristics that may influence drug selection and/or persistence, were included in the final models. A list of all covariates included in the multivariable analysis with description and format are presented in Table 13.                                                  38  The value ‘lower’ was also assigned for patients when the index prescriber did not prescribed any course of TNFα antagonist in the cohort in the preceding year.    133 Figure 13: Calculating the proportion used to assign the value of PPD                   Identify the drug initiated in these one-year courses Calculate proportion from the one-year data: number of courses with index drug divided by the total number of courses For each patient, identify the index drug, index date and the index prescriber Identify all courses initiated by the index prescriber in the year preceding the index date  Course of interest 1 year n=2 n=1 n=6 Infliximab Adalimumab Etanercept Proportion of infliximab %6.66 126 6    Courses initiated by the index prescriber time    134 Table 13: List of covariates included in the final models Variable Description Demographics  Sex  Bernoulli variable, reference=female Age at index date Four mutually exclusive categories: 18-29 years, 30-69 (reference 39 ), 70-79 and ≥80 years old The annual deductible for prescription cost at index date PharmaCare deductible level is based on annual income [141]. Six mutually exclusive categories: annual deductible of 0 dollars (reference), deductible of 1-500, deductible of 501-2250, deductible of >2250, other PharmaCare programs (no deductible level available) and no PharmaCare program Geographical area at index date Geographical area was determined based on the first three letters/digits of the patient’s postal code of residence. Five mutually exclusive categories based on: Greater Vancouver (reference), Greater Victoria, Vancouver Island (excluding Victoria), urban areas (excluding Vancouver and Vancouver Island)  and rural areas (excluding Vancouver Island) Clinical status  Physicians encounters in the year preceding the index date Continuous, one unit = one visit Inpatients admissions in the year preceding the index date Four mutually exclusive categories based on the number of admissions: no admission (reference), one admission, two admissions and at least three admissions Comorbidities (presence and severity) during the three years preceding the index date Charlson comorbidity score [176] was determined using Quan’s ICD- 9-CM and ICD-10 coding algorithm for administrative databases [175], excluding rheumatic diseases. At least two outpatient or one inpatient encounter with the diagnosis, were required.  Four mutually exclusive categories: score of 0 (reference), score of 1, score of  two and score above two   Disease duration Continuous, one unit=one year, measured from the first diagnosis of RA (inpatient or outpatient) in the data.  The presence of extra-articular manifestation during the three years preceding the index date Bernoulli, reference=no, based on recorded at least one diagnoses (inpatient or outpatient). A list of codes is presented in Appendix B, Table 28, page 271.                                                  39  Patients within age range 30-70 years were found to have a similar hazard ratio in preliminary analysis.    135 Variable Description Drug therapies  Concomitant MTX during 200 days 40   preceding the index date Bernoulli, reference = no, based on dispensing claims  Dispensing claims for NSAIDs during the year  preceding the index date Bernoulli, reference = no  Number of different antirheumatic drugs dispensed in the three years preceding the index date Dispensing claims of 10 drugs were included: MTX, hydroxychloroquine, sulfasalazine, leflunomide, azathioprine, minocycline, penicillamine, sodium aurothiomalate, prednisone, and intra-articular triamanolone or methylprednisolon.  Four mutually exclusive categories: no drug, 1-2 different drugs (reference), 3-6 different drugs, >6 different drugs  Other  Prescriber propensity for treatment discontinuation An individual Bernoulli variable for the index prescriber (58 prescribers) Calendar year at index date The variable allowed controlling for secular trends in clinical practice [157-160] and availability of drugs 41 .  Eight yearly categories were included for the years 2001-2008 (reference = year 2001). ICD-10- International Classification of Diseases, 10 th  edition; ICD-9-CM - International Classification of Diseases, 9 th  edition, clinical modification; MTX – methotrexate; NSAID – nonsteroidal anti-inflammatory drug                                                  40  The time interval (200 days) chosen was based on mean plus two standard deviations of between- dispensing intervals of MTX in the study cohort 41  Adalimumab gain entry to the Canadian market at the end of 2004, while the infliximab and etanercept were available from 2001.    136 5.3.3 Outcome  The outcomes variables were defined and assigned values consistent with those used in the analysis presented in Chapter 3 (Section 3.3.1, page 60). Specifically, the two outcomes variables were (a) a continuous persistence variable and (b) a binary censor variable. Drug persistence in years was calculated as the difference between the date of drug discontinuation and the index date. Drug discontinuation was ascertained by either switching to another antirheumatic monoclonal antibody or immunomodulator (infliximab, adalimumab, etanercept, anakinra, rituximab, abatacept, certolizumab and golimumab), or the elapsing of a drug-free interval of 180 days. A drug-free interval was defined as a period without additional dispensing claim of the same pharmaceutical component after the days-supply of the latest refill was exhausted (Chapter 2, Figure 5, page 22). The date of discontinuation was set to the end of the days-supply of the last refill before discontinuation or the date of the first dispensing of a different monoclonal antibody or immunomodulator, whichever was earliest. Patients were considered censored if they were continuously treated with the first TNFα antagonist up to December 31, 2009 (end of follow-up period) or when their date of discontinuation occurred during an interruption of more than six days in the provincial MSP coverage. Death and emigration from the province are the most common causes of interruptions in the provincial MSP coverage. For details on assigning the values of the outcome variables refer to Appendix B, Figure 24, page 271. Original data on the number of days-supply recorded in PharmaNet were unreliable (Chapter 2, Section 2.2.4.1, page 47); hence, we also calculated the expected number of days-supply based on the estimated number of vials dispensed, which was imputed using the cost field in    137 the dispensing record. We used the longest duration of days-supply, recorded or calculated, to determine both the length of drug-free interval and the date of drug discontinuation. For further discussion on calculating the number of days-supply refer to Chapter 2, Section 2.2.4.1, page 47.  5.3.4 Statistical Analysis  Summary statistics of baseline characteristics were compiled across the three drug groups. We assumed normality of the continuous variables compatible with large sample size (>2500) and the central limit theorem. The significance of differences between the drug groups was assessed with one-way analysis of variance (ANOVA) F-test for continuous variables and the Pearson’s Chi-square test for categorical variables. All statistical tests were two-sided. All calculations were performed using the SAS software package [178].  The product-limit method and the log rank test were used to estimate and compare persistence across drugs and levels of PPD. The crude and adjusted hazard ratios for drug discontinuation were estimated using two different approaches in a Cox proportional hazards regression. In the first approach, an analysis of nonclustered data, we ignored any possible correlation between patients who received a prescription for TNFα antagonist by the same prescriber. The model included an individual Bernoulli variable for each prescriber to allow individualization of the risk of discontinuation by prescriber (prescriber propensity for treatment discontinuation). In the second approach, a marginal model of clustered data, we adjusted for possible correlation between patients who received a prescription for TNFα antagonist by the same prescriber,    138 assuming that patients treated by the same prescriber were more similar to each other than to patients treated by other prescribers. We used Lee, Wei and Amato’s approach to clustered survival data [261]42  using PHREG procedure in SAS [262,263]. The dependence structure between failure times was not specified in the model; however, possible correlation was accounted for in the variance-covariance matrix by the use of a robust sandwich estimator. We chose this method based on the assumption that PPD was not normally distributed, as was shown in a study examining physician preference to aggressive treatment by patient’s age [264]. Under the circumstances of nonnormality of PPD, we could not specify a distribution for the random effect. The results were presented as hazard ratios for the effect of higher versus lower PPD, with 95% confidence intervals (CIs) and p-values. Two-tailed p-values smaller than 0.05 were considered to be statistically significant. The proportional hazards assumption was assessed by testing for the significance of interactions between the variables and time or the natural logarithm of time, as well as by plotting the scaled Schoenfeld residuals [177]. The validity of the linearity assumption of continuous variables was assessed by log likelihood tests in models that included either categorical or polynomial variables. If nonlinearity was detected, the variable was categorized. Interaction between drug and PPD was planned a priori, but was included in the final models only if was statistically significant (a two-tailed p-value < 0.1).                                                   42  Discussion of this method is beyond the scope of this dissertation, and can be found in the dissertation by Bo 2008 [482].    139 The possible confounding effect of PPD (using the three thresholds) on comparative persistence43 was estimated. It was defined as increasing or decreasing the crude adjusted hazard ratios (or their 95% confidence limits) by at least five percentages (similar to the definition in the study presented in Chapter 3, Section 3.3.2, page 62). 5.3.5 Analysis of Switchers  We analyzed patients who switched to a second course of TNFα antagonists within 180 days after discontinuing a first course. Patients were divided into two groups, based on the level of PPD assigned to the TNFα antagonists prescribed as their in the first course. We compared the proportions of patients who were switched to a second drug with higher a PPD. The decision to select a second drug was conditional on failure of first drug; hence determining the value of PPD for the drug prescribed in a second course based on data pertaining to first courses prescribing only could not be considered valid. Accordingly, PPD for the second course was calculated using all courses of TNFα antagonists initiated by the prescriber to all RA patients                                                  43  Comparative persistence was defined and estimated as presented in Chapter 3, Section 3.3.2, page 3 and Chapter 4, Section 4.3.5, page 3. We included three pairwise comparisons for hazard ratio of discontinuing (a) infliximab compared to etanercept, (b) infliximab compared to adalimumab and (c) adalimumab compared to etanercept.    140 in the study cohort44 in the previous year. The threshold between assigned levels was set to 60%, consistent with analysis of the first course. We duplicated the analysis for switchers who were repeatedly treated by the same prescriber. 5.4 Results 5.4.1 Study Cohort and Baseline Characteristics The study cohort included 2,742 RA patients treated by 58 prescribers who treated at least five patients45 with infliximab (46 prescribers), adalimumab (49 prescribers) and/or etanercept (58 prescribers). Etanercept was the most frequently prescribed TNFα antagonist (1718 patients, 63%), versus infliximab (571 patients, 21%) versus adalimumab (453 patients, 16%). More patients on etanercept (70%) were treated by a drug with higher PPD, compared to infliximab (34%) or adalimumab (19%). Baseline characteristics are presented in Table 14. Patients                                                  44  We examined courses for 2,923 RA patients who initiated TNFα antagonists by December 31, 2008 (the cohort presented in Chapter 3, Section 3.4, page 3). We did not exclude patients in which the first course was prescribed by ‘low volume’ prescriber, since the second course (or the third) could be initiated by prescriber included in the cohort analyzed in this chapter. 45  A total of 2,923 patients that were included in RA cohort analyzed in Chapter 3 (Section 3.4, page 3) and Chapter 4 (Section 4.4.1, page 3). We excluded 181 patients that were prescribed by physician with initiated less than five courses in the RA cohort.    141 treated with adalimumab were significantly older and had a lower income (lower annual deductible level for prescribing cost). They also had the lowest prevalence of dispensing claims for nonsteroidal anti-inflammatory drugs (NSAIDs) and lowest rates of admission to hospital in the year prior to initiating the study drugs. Residents of the Greater Vancouver area were more commonly treated with infliximab. Patients who received infliximab also had the highest prevalence of concomitant MTX therapy versus patients on etanercept with the lowest.  5.4.2 Effect of Prescriber Preference on Persistence The product limit plots for the drug persistence are presented by specific TNFα antagonist (Figure 14A’ page 143) and by PPD level (Figure 14B, page 144). Similar persistence was observed with the three TNFα antagonists (median persistence with infliximab 3.88 years [95% CI 2.96-5.25], adalimumab 3.33 [2.63-4.10] and etanercept 3.87 [3.36-4.40], log rank test p- value=0.15). These results are consistent with the analysis presented in Chapter 3 (Section 3.4.1, page 64). Higher PPD was associated with improved persistence. The median persistence of patients treated with a drug with higher PPD was 4.28 (95% CI 3.70-4.90) compared to persistence with lower PPD of 3.27 years (95% CI 2.84-3.84), log rank test p-value=0.017. The crude and adjusted hazard ratios for higher PPD, with and without accounting for clustering by prescribers, are presented in Table 15 (page 24, column: PPD threshold 60%). Higher PPD was associated with a significant decrease of 14-15% in the adjusted hazard for    142 Table 14: Baseline characteristics Variable Infliximab (n=571, 21%) Adalimumab (n=453, 16%) Etanercept (n=1718, 63%) P-value for comparison Courses with higher PPD n (precentages from courses of this drug) 193 (34%8) 84 (19%)  1198 (70%) <0.0001 Demographics     Females, n (%)  403 (71%)  326 (72%) 1239 (72%)   NSS Age at index (years) median (range) 56 (18-87) 58 (22-91) 56 (18-92) 0.003 Annual deductible for prescription cost n (%)       None 47 (8.3%) 80 (18%)  199 (12%)  <0.0001   $1-500 33 (5.8%) 53 (12%)  152 (8.9%) 0.004   $501-2250 92 (16%)  109 (24%)  315 (18%) 0.004   >$2250 35 (6.1%) 38 (8.4%) 143 (8.3%) NSS Residence in Greater Vancouver and Victoria n (%) 341 (60%)  224 (50%)  782 (46%) <0.0001 Clinical status     Physician visits median (range) 33 (3-158) 31 (2-112) 32 (3-136) NSS At least one admission to hospital n (%) 104 (18%) 63 (14%) 340 (20%) 0.01 Extra-articular manifestations n (%) 28 (4.9%) 14 (3.1%) 60 (3.5%) NSS Presence of comorbidity (Charlson comorbidity score > 0)  n (%) 113 (20%) 95 (21%) 383 (22%) NSS RA disease duration (years) median (range) 9.2 (0.1-17.9) 7.7 (0.3-17.9) 8.0 (0-17.8) NSS RA drugs     Concomitant MTX n (%) 388 (68%) 264 (58%)  856 (50%) <0.0001 Dispensing claims for NSAIDs n (%) 307 (54%) 214 (47%)  923 (54%) 0.04 Number of different antirheumatic drugs median (range) 4 (0-8) 4 (0-8) 4 (0-9) NSS %- percent; $-Canadian dollars; MTX-methotrexate; n-number of patients; NSAID- nonsteroidal anti- inflammatory drug; NSS – not statistically significant; PPD – prescriber preference for the prescribed drug;     143  Figure 14: The product limit estimates of persistence  A. Drug persistence by drug     144 B. Drug persistence by PPD level    145 discontinuation compared to lower PPD. No significant interaction between the drug and PPD was observed in either model. The results of analyses with thresholds of 60%, 70% and 80% for PPD were robust (Table 15 and Appendix D, Figure 26, page 282). Higher PPD was associated with a significantly decreased hazard for discontinuation by 12-15% in the three multivariable models accounting for clustering (marginal models). The adjusted hazard ratio  Table 15: Hazard ratios for drug discontinuation, higher PPD versus lower PPD Approach Hazard ratio (95% CI), p-value  PPD threshold 60% PPD threshold 70% PPD threshold 80% Patients with higher PPD (proportion) 0.52 0.45 0.36 Nonclustered data analysis Crude 0.88 (0.79-0.98), 0.02 0.88 (0.80-0.99), 0.03 0.89 (0.80-1.00), 0.04 Adjusted 46  0.86 (0.76-0.98), 0.03 0.88 (0.77-1.00), 0.06 0.87 (0.75-1.00), 0.06 Marginal modeling of clustered data Crude 0.88 (0.78-1.01), 0.05 0.88 (0.78-1.01), 0.07 0.89 (0.79-1.01), 0.06 Adjusted 47  0.85 (0.76-0.96), 0.01 0.87 (0.78-0.97), 0.01 0.88 (0.82-0.98), 0.02 CI – confidence interval; PPD – prescriber preference for the prescribed drug                                                  46  Adjusted by drug, calendar year, patient’s demographics and clinical status, and prescriber   47  Adjusted by drug, calendar year and patient’s demographics and clinical status     146 based on the nonclustered data analysis did not reach the significant level for thresholds of 70% and 80%, perhaps as a result of decreasing numbers of patients treated with a drug of higher PPD.   5.4.3 Effect of Prescriber Preference on Comparative Persistence We estimated a possible confounder effect of PPD on comparative persistence by comparing estimates of hazard ratios for discontinuation using unadjusted modeling (crude estimated of hazard ratios, using nonclustered and clustered data approaches) and models that included one additional covariate – PPD. We used three thresholds to assign the value of PPD (as described in Section 5.3.2, page 132). The results are presented in Table 16. We found that PPD did have possible confounder effects on hazard ratios of infliximab compared to adalimumab in four of the six analyses, including all three models for clustered data. A possible confounder effect of PPD on hazard ratios comparing infliximab and etanercept was also demonstrated in one of the models of six analyses (threshold of 60% for the value of PPD, clustered data). No confounder effect was demonstrated for six analyses yielding hazard ratios for discontinuing adalimumab compared to etanercept.     147 Table 16: Possible confounder effect of PPD on comparative persistence Model Hazard ratio (95% CI)  Infliximab versus etanercept  Infliximab versus adalimumab Adalimumab versus etanercept Nonclustered data analysis    No adjustment (crude estimate) 0.97 (0.84-1.13) 1.04 (0.88-1.22) 0.94 (0.77-1.15) Adjustment for PPD (threshold=60%)  0.98 (0.83-1.17)  Adjustment for PPD (threshold=70%)    Adjustment for PPD (threshold=80%)    Marginal modeling of clustered data    No adjustment (crude estimate) 0.99 (0.82-1.18) 1.09 (0.92-1.29) 0.91 (0.71-1.17) Adjustment for PPD (threshold=60%) 0.93 (0.78-1.11) 1.01 (0.85-1.21)  Adjustment for PPD (threshold=70%)  1.03 (0.87-1.22)  Adjustment for PPD (threshold=80%)  1.04 (0.89-1.22)  Presented only for models in which PPD changed the magnitude of point estimate (or confidence limits) of the hazard ratios for discontinuation comparing different drugs by at least 5% CI – confidence interval; PPD – prescriber preference for the prescribed drug 5.4.1 Analysis of Switchers The proportions of patients who switched to a second drug with higher PPD are presented in Table 17. In patients who were repeatedly treated by the same physician (N=500), the likelihood of switching to second drug with higher PPD was increased when discontinuing a drug with lower PPD compared to discontinuing a drug with a higher PPD, 19.7% versus 11.1% (Chi-square test p-value = 0.008).    148 Table 17: Analysis of switchers  Second course prescribed by any physician nN=603) First and second course prescribed by the same physician (n=500)  n Second course with higher PPD n (%) n Second course with higher PPD n (%) Group 1: First course of a TNFα antagonist with lower PPD 282 62 (22.0) 239 47 (19.7) Group 2: First course of a TNFα antagonist with higher PPD 321 51 (15.9) 261 29 (11.1) Chi-square test comparing groups 1 and 2 NSS p=0.008 %- percent; n-number of patients in the category; NSS – not statistically significant; PPD – prescriber preference for the prescribed drug; TNFα – tumour necrosis factor alpha 5.5 Discussion We investigated the possible effect of prescriber preference for a specific TNFα antagonist (PPD) on two prescribing situations: discontinuation of a first course and drug selection for a second course. In the first situation, that of discontinuing a first TNFα antagonist course, the results support our hypothesis that higher PPD is associated with improved TNFα antagonist persistence. We also demonstrated that PPD is a possible confounder of the hazard ratios for discontinuation. The second situation that we studied was drug selection48 in patients who initiated a second course of TNFα antagonists (switchers). In switchers with continued                                                  48  We use the term selection to describe the decision on which of the three available TNFα antagonists to use. This decision is conditional on the decision to prescribe a TNFα antagonist.    149 prescribing by the same prescriber, we demonstrate an increased likelihood of switching to a second TNFα antagonist with a higher PPD when patients discontinued a first course of TNFα antagonist with a lower PPD.  5.5.1 Prescribing Habits and Prescribing decisions Prescribing decisions are complex processes influenced by multiple factors. Prescribing decisions can be divided into decisions on treatment initiation, adjustment and discontinuation. The effect of the prescribing habits was previously studied in RA treatment initiation and discontinuation.  5.5.1.1 Treatment Initiation  Treatment initiation can be described as a sequel of three decisions: (a) whether to initiate any treatment (default- no treatment), (b) the choice of a therapeutic class and (c) the selection of an individual drug from within the therapeutic class. Each of these decisions is conditional on the previous one. We do not discuss the decision to initiate RA treatment, because TNFα antagonists are rarely indicated as first line therapy (Appendix A, Section A.1.3, page 218); hence, the decision to treat the patient can be assumed to have already been made.  Patients who required treatment augmentation due to decreased benefit with a first line drug (synthetic antirheumatic drugs), could be treated with either synthetic antirheumatic drugs (a different drug or more commonly - combinations therapy) or a TNFα antagonist [265-267]. The decision whether to prescribe synthetic antirheumatic drugs or switch to a TNFα    150 antagonist is often motivated by disease activity and prognostic factors [267,268] but may also be influenced by the cost of the drug [268-270]. Prescriber preference49 was found to be the most important determinant in initiating treatment with TNFα antagonists (compared to prescribing synthetic antirheumatic drugs) in an on-line survey by Cush et al 2005 [271]. Of 1,023 United State (U.S.) rheumatologists, 48% selected provider preference versus patient preference or payer guidelines (roughly 20% each).  The role of prescriber preference for the therapeutic class50, based on previous TNFα antagonist prescribing patterns, was also demonstrated by Curtis et al 2010 [272,273]. They analyzed the demographic and clinical data of patients with RA enrolled in the Consortium of Rheumatology Researchers of North America (CORRONA)51 cohort study. Prescriber preference for the therapeutic class was determined based on the proportion of initiators (or prevalent users) of TNFα antagonists compared to all initiators (or prevalent users) of either TNFα antagonists or synthetic antirheumatic drugs in patients populations treated by same                                                  49 In the study by Cush et al [271], the term prescriber preference describes whether the physician favor initiating a TNFα antagonist over other treatment options, such as switching to or adding-on a synthetic antirheumatic drug. The use is similar to the term prescriber preference for the prescribed drug used in this thesis, except that is refers to the therapeutic class and not to a specific drug. 50  In the study by Curtis et al [272,273], the term prescriber preference describes whether the physician favor initiating a TNFα antagonist over other treatment options, such as switching to or adding-on a synthetic antirheumatic drug. The use is similar to the term prescriber preference for the prescribed drug used in this thesis, except that is refers to the therapeutic class and not to a specific drug. 51 The CORRONA is a longitudinal cohort that includes patients with RA, psoriatic arthritis, osteoarthritis, and osteoporosis treated by rheumatologists in 35 states in the U.S., founded in 2000. Clinical data is collected prospectively.     151 prescriber in the preceding year52. The analysis was conducted using generalized estimating equations to account for the clustering of patients by prescriber. The results of the study indicated that higher preference was associated with increased likelihood of initiating treatment with a TNFα antagonist53, especially in MTX naïve patients and in patients with low or moderate disease activity. Nevertheless, the prescriber preference for the therapeutic class only slightly improved the model’s discrimination54.  The role of prescriber preference for a therapeutic class was also studied in initiation of  selective inhibitors of cyclooxygenase-2 (COX-2) compared to nonselective NSAIDs in patients with RA where it was demonstrated to be the most important factor in this decision55 [274-276]. Similar association was observed in the decision to prescribe either generic or nongeneric serotonin reuptake inhibitors and serotonin- norepinephrine reuptake inhibitors [277].                                                  52  Preference was determined two ways: using data on initiators or data on prevalent users. The initiator proportion, for example, was calculated as (number of initiators of a TNFα antagonist)/(number of initiators of either a TNFα antagonist or a synthetic antirheumatic drug), data used in the numerator and denominator were on patients prescribed by the same prescriber in the previous year. For calculating the proportion of prevalent drug users, the researchers used data on all drug users, and not only initiators. 53  The highest and middle tertiles of prescriber preference were associated with a 1.61-4.89 and 1.14-2.44 greater likelihood of initiating TNFα antagonists compared to the lowest tertile, respectively, after adjusting for demographic and clinical factors. 54  The c-statistics increased from 0.79 in a model that included only demographic and clinical variable to 0.81 in a model that also included prescriber preference. 55  Based on standard R-square as a measure of prediction [274,275], C-statistic [275] and Bayesian modeling [276]).    152 The next prescribing decision is the selection of a specific TNFα antagonist. It is accepted that selection of a specific TNFα antagonist depends on prescriber and/or patient preference, since the benefit and harm profiles of the three drugs are considered to be similar [278-280] despite limited relative effectiveness data. Prescriber and patient preferences for an individual TNFα antagonist are likely to be correlated for the following reasons. First, patient preference was found to be strongly correlated with prescriber preference in studies of NSAID treatment of RA patients [281,282], although this was not studied in patients treated with TNFα antagonists. Second, 67% of 77 Canadian rheumatologists surveyed reported concordance with the patient preference more than 80% of the time56 [283].  Patient preference for route of administration of TNFα antagonist (Appendix A, Section A.3.6.1, page 262), dosing schedule [284,285] or the specific TNFα antagonists [140,286] has been studied. Patients reported preferring etanercept to methotrexate, gold and leflunomide[286] or adalimumab over infliximab and etanercept [140].  Physician preference for the individual TNFα antagonists was studied by Kamal et al 2006 [268]. In respond to mailed questionnaires, 68% of 428 U.S. rheumatologists who responded said they preferred etanercept over adalimumab (preferred by 28%) or infliximab (20%) and considered etanercept the most efficacious of the three drugs with less perceived harm [268].                                                   56  Patients were not surveyed separately, but rather the results presented are from the rheumatologists reporting.    153 We are unfamiliar with any study examining the association between prescriber preference for a particular drug within a therapeutic class and drug selection, in RA patients or other diseases, although a correlation similar to the one observed regarding therapeutic class above could be reasonably expected.  5.5.1.2 Treatment Adjustment Treatment adjustment consists of dose adjustment of the drug in use (Appendix A, Section A.3.6.2, page 265) as well as add-on therapies of additional drugs from the same or different therapeutic class or different class. In decision to adjust treatment in RA patients, physician have been found to be more influenced by objective factors, such as the disease activity score while patients are more influenced by subjective factors, including function, satisfaction with current treatment and motivation to get better [287,288]. The effect of prescribing habits on treatment adjustment strategy has not, to our knowledge, been studied. 5.5.1.3 Drug Discontinuation Potentially, two aspects of prescribing habits could induce differences in discontinuation decisions: (a) between-physicians differences in response to similar clinical situations (prescriber propensity for treatment discontinuation) and (b) different response to similar clinical situations by the same physician, a response that could be corresponded to the prescriber preference for a particular drug (the rationale is presented in Section 0, page 159). Between-physicians differences – The main reasons for discontinuing TNFα antagonists in patients with RA, as recorded by the care-providing physicians, are decreased benefit and perceived harm (Appendix A, Section 0, page 255). However, physicians may respond    154 differently to similar clinical situations, a pattern that can be described as prescriber propensity for treatment discontinuation. Differences in response to harm were reported by Cush et al 2005 [271]. American rheumatologists were asked in an on-line survey whether they would discontinue TNFα antagonists’ treatment in response to a number of harmful effects. Higher agreement was observed in the events of severe infection (98% of the rheumatologists reported they would discontinue TNFα antagonist treatment in the event of septic arthritis, 97% in pneumonia and 91% in cellulitis) and worsening of congestive heart failure (91%) compared to other infections (50-80%) and following hospitalization (55%). Disagreement on the response to harmful events represents differences in prescriber propensities to treatment discontinuation. Differences in prescriber response to decreased benefit with TNFα antagonists therapy was recently reported by Zhang et al 2011 [48,289]. Investigators examined the effect of prescriber propensity for treatment discontinuation (they used the term ‘prescriber preference’)57 on TNFα antagonist discontinuation due to decreased benefit within the first year of treatment in the CORRONA RA patient cohort. Prescriber propensity for treatment discontinuation was modeled by clustering patients at prescriber level using mixed effect logistic regression model with a random intercept. The study demonstrated the importance of clustering by prescribers, even after adjustment for both baseline disease activity and improvement in disease activity.                                                  57  In the study by Zhang et al [48,289], the term prescriber preference describes the degree to which the physician favors discontinuation of the TNFα antagonist treatment. In this dissertation, we used the term prescriber propensity for treatment discontinuation to describe this phenomenon.     155 The authors concluded: “We observed a significant clustering effect by physician associated with switching58, consistent with the idea that there are individual physician-specific factors that affect treatment choices”. These factors may include prescriber propensity for treatment discontinuation, the patient’s insurance status (relevant to the U.S. health system), the formulary status of the particular TNFα antagonist or the prescriber’s case mix of RA patients.  Differences in response of the same physician – Our study was the first to report an effect of PPD (prescriber preference for an individual TNFα antagonist, or any other drug) on discontinuation in RA patients (or any other treatment population). Our model included both a variable for PPD and adjustment for prescriber preference for discontinuation.  Generally, several approaches are appropriate model prescribing habits (preference for the prescribed drug/therapeutic class and propensity to discontinuation) in regression analysis, and they are described in Section 5.5.2, page 156.                                                  58  In the study by Zhang et al, the term switching included “either discontinuation of the anti-TNF agent and/or changing to a different agent for reasons other than safety/tolerability” [48].    156 5.5.2 Effect of Prescribing Preference The effect of prescriber preference for the drug/therapeutic class and prescriber propensity for treatment discontinuation on prescribing decisions had previously been studied using administrative (claim) data, but not clinical data or physician survey. We identified two sets of published analyses of administrative (claim) data that had modeled prescribing habits and their effect on prescribing decisions. In the first set of studies, studies of treatment initiation, prescriber preference for a particular drug/therapeutic class was modeled [272-277]. The second set of studies modeled prescriber propensity for treatment discontinuation (describing the degree to which the prescriber favored drug discontinuation over other alternatives) and included a single study by Zhang et al 2011 [48,289]. Our study was the first to include both types of prescriber variables (prescriber preference for the prescribed drug and prescriber propensity for treatment discontinuation) in one model.  Analyses of administrative data for the effect of prescribing habits can be performed using nonclustered or clustered data approaches. The difference in the outcome variable between the abovementioned two sets of studies leads to discrepancies in interpretability of the results between studies of treatment initiation versus studies of discontinuation (Figure 15 and Figure 16). Below, we present the methods and the interpretability of results of modeling the effect of prescribing habits on drug discontinuation.    157 Figure 15: Adjusting for physician preference in studies of treatment initiation    Prescribing situation Initiation Discontinuation Non-clustered data Clustered data Analytic approach Variables for prescribers  Variable for drug preference (PPD) Mixed effect Marginal  Adjustment for prescriber preference for a particular drug/therapeutic class    158 Figure 16: Adjusting for physician preference in studies of drug discontinuation   Prescribing situation Initiation Discontinuation Non-clustered data Clustered data  Analytic approach Variables for prescribers  Variable for drug preference (PPD) Mixed effect Marginal  Adjustment for prescriber propensity for treatment discontinuation  Adjustment for prescriber preference for a particular drug/therapeutic class    159 In analysis of nonclustered data, there are two acceptable approaches: (a) modeling the individual prescriber propensity for treatment discontinuation and (b) modeling prescriber preference for a particular drug/therapeutic class at treatment initiation (e.g., PPD). In the first approach, the analyst includes a Bernoulli variable for each individual prescriber (Chapter 3, Section 3.3.2, page 62 and Chapter 4, Section 4.3.5, page 92). This allows individualization of the likelihood of discontinuation (prescriber propensity for treatment discontinuation). Using this approach, we assume that the effect of all other predictors of discontinuation is uniform across all prescribers. The second approach for nonclustered data analysis is to include a variable that describes preference to initiate a particular therapeutic class or an individual drug (similar to the variable PPD used in this chapter, Section 5.3.2, page 132). Prescriber preference for the prescribed drug/therapeutic class is assigned based on previous prescribing habits of the prescriber, using either initiating data (this study and Curtis et al 2010 [272]59), or all prescribing data (prevalent users  [272,277]59,60). The assumption that previous prescribing accurately describes prescriber preference for the prescribed drug was validated by Mark et al 2004 [290], who compared stated preference for various antialcoholism drugs measured using validated questionnaires with prescribing data based on actual prescribing decisions.                                                   59  The studies cited assessed treatment initiating. No study assessing discontinuation adjusted for preference for the prescribed drug or therapeutic class.  60  The study by Baser [277] modeled prescriber preference for a particular therapeutic class (selective serotonin reuptake inhibitors versus serotonin- norepinephrine reuptake inhibitors) and test for its strengths as an instrumental variable.     160 Alternatively, analysis of clustered data (by care-providing physician) can account for prescriber propensity for treatment discontinuation compared to other treatment alternatives. Investigators can use either mixed effect models [48,289] or marginal models (similar to the modeling presented in this chapter, Section 5.3.4, page 137) (Figure 16). Analysis of clustered data is based on the assumption that patients treated by the same prescriber were more similar to each other (and therefore had similar likelihood to discontinue) than to patients treated by other prescribers. One of the reasons for clustering effect is distinct prescriber propensity for treatment discontinuation, although other reasons are possible as well. 5.5.3 Effect of Prescriber Preference for the Prescribed Drug on TNFα Antagonists Persistence  We suggest three possible explanations for the finding of improved persistence with increased prescriber preference for the prescribed TNFα antagonist (PPD). First, the results could be explained in light of the theory of cognitive dissonance [291]. In line with this theory, treating with a drug believed to be inferior (lower PPD) induced dissonance61. In indefinite clinical situations, such as mild harmful effect, or questionable benefit, early drug discontinuation supports the physician’s belief that the selected drug was inferior compared to the alternatives.                                                  61  Dissonance is the uncomfortable feeling caused by holding conflicting ideas simultaneously.    161 On the other hand, in similar indefinite clinical situations, a drug with a higher PPD might be continued, to support a belief in its superiority. The interpretation of our results using the theory of cognitive dissonance is restricted by a lack of direct measurement of dissonance. Direct measures of cognitive dissonance are complicated and involve activities beyond the scope of the current study such as questioning the physician both at the time of treatment and initiation and discontinuation [292]. Database analyses are relatively inexpensive to perform, and in many cases, the results should be interpreted as signals of a possible association that justify further research. The indirect measure of dissonance we used, PPD, could also be influenced by factors unrelated to prescriber beliefs, such as limited availability of a drug or higher relative cost. Secondly, the observed association between higher PPD and improved persistence might be the result of a correlation of both PPD and persistence with past experience with the drug, namely the total number of patients a prescriber has previously treated with the same drug. Increased experience with a specific drug may be associated with higher PPD as a result of the algorithm we used to define PPD, but also because of a tendency to continue doing what is familiar. Theoretically, increased experience with a specific drug should also be associated with improved persistence, due to improved patient selection which results in decreased risk of harmful effects or decreased benefit. Lastly, a prescriber’s stated preference for a particular drug has been shown to correlate with patient preference [281,282], although this correlation has not been studied in relation to TNFα antagonist therapy, nor has preference as estimated based on previous prescribing habits been correlated with patient preference in TNFα antagonists treatment or other treatment situations.    162 If this correlation is valid in the context in our analysis, then improved persistence may have been a result of higher patient rather than prescriber preference. The association between patient preference and persistence was discussed in Appendix A, Section A.3.6.1, page 262. Implications We have demonstrated that care-provider characteristics, specifically prescribing habits, besides patient characteristics and an individual TNFα antagonist effect, also has an impact on persistence and this finding calls into question the validity of the assumption that persistence is a valid measure that balances benefit and harm. In addition to prescriber propensity for treatment discontinuation (Chapter 3, Section 3.4.2, page 68, [48,271,289]), we have demonstrated that prescriber preference for a particular TNFα antagonist also affects discontinuation. We therefore conclude that clinical and policy decisions based on the finding of persistence studies as a measure of effectiveness and harm may be unsound.   Awareness of the possible role of the cognitive dissonance in clinical decision-making based on beliefs unsupported by clinical evidence can contribute to the development of education programs for prescribers, as well as health policy development. We highlight a situation in which prescriber preference for a particular TNFα antagonist was not based on clinical evidence as confirmed by inconsistencies in treatment patterns amongst prescribers. Cognitive dissonance may also play a role in improving adherence to clinical or health policy guidelines in that insights gained about actual prescribing behaviour in the absence of rationale can be used to motivate rational prescribing. Most of this type of research is warranted to identify and explore similar situation for the purpose of developing evidence-based clinical guidelines and identifying effective policy implementation approaches.     163 Future research We suggest three alternative explanations for our findings, which could lead to further research in the corresponding direction of: (a) directly measuring cognitive dissonance, (b) analyzing the effect of prescriber experience on persistence and (c) measuring patient preference and its correlation with prescriber preference. In addition, we propose to directly measure prescriber preference for individual TNFα antagonist drugs, by using questionnaires at the time of treatment initiation and drug discontinuation. The correlation between preference and dissonance, preference and previous prescribing habits and preference and persistence could thereby be assessed. We also suggest extending the research to study the effect of prescriber preference in pharmacological treatments of chronic diseases, such as Alzheimer disease, chronic lung disease and other rheumatologic diseases, to test for the generalizability of our findings. 5.5.4 Effect of Prescriber Preference on Comparative Persistence with TNFα Antagonists We demonstrate that comparative persistence was confounded by prescriber preference for a particular TNFα antagonist. This means that the effect of prescriber preference on persistence may be different for each TNFα antagonist. The effect is probably more pronounced in patients treated with infliximab as only hazard ratios comparing infliximab were affected. We propose that the close monitoring of patients treated with intravenous infliximab and regular communication between patients and physicians that must of necessity occur with intravenous administration is the root of this greater influence of PPD on persistence. In drugs administered subcutaneous, the effect is probably “diluted” by patient behaviour.    164 Implications Excluding PPD from analysis of comparative persistence may have been another factor that caused heterogeneity in previous estimated of comparative persistence with TNFα antagonists (an original unpublished systematic review and meta-anaysis). As a result, findings would appear to have been biased at least in some of the studies. This provides additional support for our conclusions about the limited validity of the current literature on comparative persistence with TNFα antagonists in RA. Therefore the available research to date on the comparative persistence of TNFα antagonists is an inappropriate basis for clinical and policy decisions.     Future research We suggest that adjusting for PPD is warranted in studies of comparative persistence of TNFα antagonists in RA patients because it may confound the results. We also suggest extending the research to study its possible confounding effect in comparative persistence studies of other pharmacological treatments of chronic diseases. 5.5.5 Analysis of Switchers We suggest three possible explanations for the finding of an increased likelihood of switching to a second drug with higher PPD when patients discontinued the TNFα antagonist of lower PPD (in switchers with continuous prescribing by the same prescriber). First, this finding is also supported by the theory of cognitive dissonance because selecting a second drug believed to be superior would further decrease dissonance. Second, when a patient failed on a drug with a higher PPD, the second TNFα antagonist could be of higher PPD only if the prescriber preference had changed over time (between initiation of the first course and the second course).    165 This is probably not very common, especially with short persistence on the first and failed course. As a result, artificially low percentages of patients were treated with both first and second drug of higher PPD. Last, the results may be confounded by other factors that influence drug selection (for example age or accessibility to health services). We did not model the likelihood or adjusted for these other factors, but rather used crude proportions.   Future research While the effect of prescriber preference was often studied in the context of initiating treatment, we did not find studies examining the selection of a second drug i.e., studies of TNFα antagonist sequencing. Our attempt to study factors that influence drug sequencing failed due to insufficient sample size and this analysis is not presented in this dissertation. Therefore, this appears to be a useful area for further research.  5.5.6 Strengths and Limitations In this population-based study, we used a large sample of patients with a prolonged duration of follow up. These features increased the power of the study to detect differences in persistence. We adjusted for multiple variables that might have influenced selection of a specific TNFα antagonist and/or drug persistence. Treatment with TNFα antagonists is especially suitable for studying effects of PPD, since the available evidence suggests that the drugs are indeed comparable (Appendix A, Section A.3.1.1, page 241). Published clinical guidelines recommend using either TNFα antagonist (interchangeably) based on assumed similar effectiveness of the three drugs (Appendix A, Section A.1.3, page 218). In addition, diverse indications for discontinuation are presented in these guidelines (Appendix A, Section A.3.2,    166 page 251), and it has been demonstrated that prescribers differ in the way they interpret and respond to decreased benefit [48,289] or perceived harm [271]. There are several limitations to ascertaining PPD based on TNFα antagonist initiation in the past. First, the PPD was not defined for all patients, for example, for the first patient treated by the prescriber, or if the physician did not prescribed this class of drugs in the preceding year. Second, we assumed that increased loyalty for a particular TNFα antagonist reflects preference, but we did not measure preference directly. As mentioned, stated preference and prescribing habits were shown to correlate in previous research [290].Third, we measured PPD at the time of treatment initiation and not at discontinuation, and prescriber preference might have changed over time. Finally, we assumed strong preference for one drug only and regarded situations in which the prescriber preferred two of the TNFα antagonist to be a situation of treatment with a drug of lower PPD.  5.6 Conclusions  In RA patients treated with a first course of TNFα antagonists, higher PPD was associated with improved persistence.   PPD was also found to confound estimates of comparative persistence. Consequently, our results provide additional explanations for the heterogeneity observed in previous studies of comparative persistence with TNFα antagonists in RA patients.     167 CHAPTER 6:  CONCLUDING CHAPTER This thesis sought evidence on methodological aspects of measuring persistence, as well as the influence of different characteristics of the care-providing physician in estimates of comparative persistence with TNFα antagonists in RA The main findings of the research, their context and their implications are presented below. 6.1 There Were Similar Persistence Levels Observed With Infliximab, Adalimumab and Etanercept but Improved Compliance Only With Infliximab in RA Patients  Our analysis of British Columbia administrative data showed similar persistence with infliximab, adalimumab and etanercept in patients with RA (median of roughly 3.5 years) (Chapter 3, Section 3.4.1, page 64). Discontinuation of TNFα antagonist therapy was based on either (a) a drug-free interval after a missed refill or (b) switching to another antirheumatic monoclonal antibody or immunosuppressant drug. Persistence was measured using a long drug-free interval of 180 days to ascertain discontinuation, in order to minimize the likelihood that the same drug was reinitiated after only temporary interruption.  Patients treated with etanercept and adalimumab were more likely than patients treated with infliximab to interrupt drug use (for 30 days or more) after which they reinitiated treatment with the same drug (Chapter 4, Section 4.4.3, page 105). About half of RA patients who had experienced an interruption of 30 days in persistence reinitiated the same drug within another    168 150 days. This result is compatible with findings in other chronic treatment situations, such as osteoporosis therapy [293] and treatment with statins [185]).Temporary interruptions in drug persistence are a possible signal of poor compliance. Therefore, our finding suggests compliance is best with infliximab, perhaps due to more frequent encounters with health professionals due to intravenous administration of this drug.  Clinical and policy implications Our results show that using short drug-free intervals can lead to invalid persistence estimates. Short drug-free intervals of 30-60 days have often been used to ascertain discontinuation in previous studies including those of TNFα antagonists. Approximately third of the patients included in our study reinitiated the same drug after interruptions of 30 days in persistence. Approximately half of the patients reinitiated the same drug within 150 days after such interruptions. Therefore, short drug-free intervals of 30-60 days should not be used to determine overall drug persistence, especially in drugs with intermittent dosing schedule, such as TNFα antagonists. Improved compliance with a variety of treatments is associated with improved effectiveness and decreases in over-all direct medical costs. Our results showed that compliance among TNFα antagonists is best with infliximab. Therefore, any program developed to improve compliance with TNFα antagonists should specifically address why patients treated with etanercept and adalimumab are particularly less likely to comply. This approach will have the greatest impact in health outcomes and health care expenditures. The effect of poor compliance on the effectiveness of TNFα antagonists has not been well studied. The prolonged therapeutic effect of TNFα antagonists may result in a weak association     169 between poor compliance and drug ineffectiveness. Consequently, the clinical significance of relatively poor compliance with adalimumab and etanercept compared to infliximab cannot be readily established. Better compliance with infliximab should be addressed in cost effectiveness studies as well as in health budget planning. Decreased drug expenditures due to noncompliance may be offset by increases in other health care expenditures. Higher acquisition costs associated with intravenous administration of infliximab also need to be considered. Further research to assess the cost implications of our observed compliance estimates is also needed.  Research implications The use of a short drug-free interval of 60 days or less to ascertain TNFα antagonist discontinuation introduces a potential for poor compliance to distort estimates of persistence. This is because patients who experience a short interruption in treatment are likely to reinitiate the same drug and therefore falsely classified as discontinuers.  The use of a short drug-free interval of 60 days or less to ascertain TNFα antagonist exposure in studies of health outcomes may also lead to invalid treatment effect estimates due to exposure misclassification. Specifically, patients who are considered discontinuers using this method may actually reinitiate treatment and yet be analyzed as if they were unexposed. This may lead to misleading conclusions. For example, an increased risk of harm after discontinuation may be actually caused by reinitiating the drug and not by a rebound effect. Misclassification bias is also a concern when benefit is observed during exposure to the drug but is apparently diminished after the patient is wrongly considered to have discontinued the drug, a phenomenon that may actually be caused by ‘tolerance’ (i.e., diminished effect of the drug with longer exposure) in patients who continue to be exposed after reinitiating the drug.    170 6.2 Comparative Persistence with TNFα Antagonists was Sensitive to the Method Used to Ascertain Drug Discontinuation We studied the effect that different pre-specified ‘drug-free intervals’ had on discontinuation of TNFα antagonists in patients with RA (Chapter 4, Section 4.4.2, page 96). A longer drug-free interval was associated with improved persistence with every TNFα antagonist. An increase in the overall median persistence from 1.13 years, to 2.19, 2.87 and 3.70 years was found after applying drug-free intervals of 30, 60, 90, and 180 days, respectively. Similar persistence between the three drugs was demonstrated in multivariable-adjusted models by applying drug-free intervals of 30 and 60 days. Improved persistence was observed with infliximab compared to the other drugs after applying longer intervals of 90 and 180 day). Persistence with adalimumab and etanercept remained similar with the longer drug-free intervals. Research Implications We provide evidence that heterogeneity in estimates of comparative persistence on TNFα antagonists in RA can be explained by a lack of consensus on the drug- free interval used to determine discontinuation. Achieving consensus on how to choose the drug-free interval is essential  for establishing a chain of valid evidence whereby different studies can contribute data on a common and comparable measure of persistence on TNFα antagonists. This advice also applies more broadly, to persistence studies in other drug classes.     171 6.3 Physician Prescribing Preferences are Predictors and Possible Confounders in Comparative Persistence Studies We examined how prescribing preference influenced persistence on TNFα antagonists. Specifically, we studied two aspects: prescriber preference for the prescribed drug (PPD) and physician’s propensity to treatment discontinuation. The first habit, PPD, measures the frequency of prescribing the particular drug compared to other drugs within the same therapeutic class. Its value is unique for each course and is assigned to a patient  based on the first prescriber and drug and the period during which it was measured. Not only was higher PPD was associated with improved persistence (median persistence of 4.28 compared to 3.27 years in courses with lower preference, Chapter 5, Section 5.4.2, page 141) but PPD was also found to confound estimates of comparative persistence (especially in comparisons of infliximab and adalimumab (Chapter 5, Section 0, page 146) .  The second habit is a physician’s propensity for treatment discontinuation, i.e. an individual likelihood to recommend discontinuing a TNFα antagonist, with higher propensity associated with lower persistence of patients treated by this prescriber (Chapter 3, Section 3.4.2, page 68). These differences could reflect different “sensitivity” of the prescribers to benefit or harm or present difference in patients mix that were not adjusted for. Furthermore, a confounding influence on comparative persistence by a prescriber’s propensity for discontinuation was demonstrated (a change of at least 5% in the magnitude of hazard ratios in comparisons of infliximab versus adalimumab and of adalimumab versus etanercept). Persistence remained similar (Chapter 3, Section 3.4.3, page 71).    172 Clinical and Policy Implications Our research is the first to demonstrate an effect of prescribing habits (drug preference and propensity for discontinuation) on treatment persistence with any drug. The effects of prescribing habits has not been well. Awareness of this effect should contribute to physician educational programs – a change in physician preference could prolong drug persistence. Restricting drug coverage policies without addressing drug preference probably would not achieve the same effect.   Research Implications Our results showed that confounding effects due to prescriber preference and propensity for discontinuation provide another possible reason for heterogeneity in comparative persistence estimates across different TNFα antagonist studies. We suspect that differences in drug preference and discontinuation propensity between different prescriber populations in different studies might have also led to different estimates. Ignoring this effect in future research studies of persistence likely limit the validity and interpretability of the results. 6.4 Confounding is a Threat to Validity in Estimating Comparative Persistence Confounding is the main threat to validity of comparative persistence estimates. This source of bias should be considered in any interpretation and appraisal of comparative persistence research on TNFα antagonists in RA patients, and is a good advice for any disease where multiple treatments can be interchangeability used.    173 Three factors were found to confound estimates of comparative persistence with TNFα antagonists in RA patients: (a) concomitant use of MTX (Chapter 3, Section 3.4.3, page 71), (b) prescriber propensity for discontinuation (Chapter 3, Section 3.4.3, page 71) and (c) prescriber preference for a particular drug (Chapter 5, Section 0, page 146).  In addition, any predictor of persistence has the potential to confound estimates of comparative persistence if it is also associated with drug’s selection. We demonstrate a number of patient characteristics that are predictors of shorter persistence and therefore have the potential to confound comparative persistence studies of TNFα antagonists in RA. These include: (a) female sex (Chapter 3, Section 3.4.2, page 68); (b) age under 30 or over 70 (Chapter 3, Section 3.4.2, page 68); (c) admission to hospital in the year preceding the index date (Chapter 3, Section 3.4.2, page 68); (d) absence of concomitant use of MTX (Chapter 3, Section 3.4.2, page 68 and Appendix A, Section A.3.5, page 261); (e) recent/current use of NSAIDs within the year preceding drug imitation (Chapter 3, Section 3.4.2, page 68) and (f) exposure to an increased number of synthetic antirheumatic drugs (Chapter 3, Section 3.4.2, page 68 and Appendix A, Section A.3.5, page 261). Additionally, we demonstrate that prescriber propensity for treatment discontinuation (Chapter 3, Section 3.4.2, page 68) and prescriber preference for the prescribed drug (Chapter 5, Section 5.4.2, page 141) influenced persistence.    174 Previous research has not shown that disease severity is a predictor of persistence (Appendix A, Section A.3.5, page 261). This is surprising, given that patients with more severe disease will presumably experience larger improvement in symptoms (regression to the mean62) and therefore be more likely to persist on their treatment. We could not confirm or refute this preassumption in our research because we lacked clinical data on disease severity. Research Implications Failure to adjust a TNFα antagonist analysis for concomitant MTX, prescriber propensity for treatment discontinuation, and prescriber preference for the prescribed drug is a potential threat to study validity, since we have shown in our data that those variables are confounders. Therefore, future studies should include an adjustment for these factors in testing for hypothesis.  A second group of variables are predictors of drug persistence that have the potential to bias comparative persistence study results if unequally distributed between drug groups. Therefore, future studies could (a) test if these factors are confounders by testing for equal distribution of these factors between drug groups and adjusting for those factors if required; and (b) testing for an independent association between those factors and persistence.                                                  62  Regression to the mean is “the tendency of an individual at the extreme to have values nearer to the mean on repeated measurement” [1].     175 Sex, age, hospital admissions, concomitant use of MTX, use of NSAIDs, number of previous synthetic antirheumatic drugs and prescriber preference for the prescribed drug are all predictors of persistence and therefore should be included in models predicting persistence. Adjustment for prescriber propensity for treatment discontinuation could be performed in these models using a marginal approach. 6.5 Interpretation of Persistence and Comparative Persistence as Measures of Benfit-Harm Balance or Relative Benfit-Harm Balance With TNFα Antagonists in RA Patients Treatment persistence is often used as an indirect measure that integrates benefit and harm [45] in studies of RA patients. Accordingly, comparative persistence, a relative measure, has also been suggested as an indirect measure of the relative benefit-harm balance of individual drugs. In Chapter 4 (Section 4.5.3, page 117), three arguments were put forward to challenge this assumption. First, when data from previous studies were pooled, we found that factors unrelated to drug benefit and harm were given as the reason for discontinuation for about one quarter of patients. Furthermore, these reasons were not evenly distributed between groups treated with different drugs which may result in differential influence of factors unrelated to drug effectiveness on persistence. Therefore, persistence would inaccurately reflect benefit- harm balance. Second, alternative clinical approaches have been used in patients who experienced either a lack of benefit or a harmful event. Pooled data from previous studies demonstrated that between-group differences in, for example, dose adjustment were more common in patients treated with infliximab compared to adalimumab or etanercept. Lastly, we    176 demonstrated that drug properties, such as therapeutic effect and harm, and patient characteristics were not the only factors that determine persistence. Persistence was also influenced by prescriber propensity for treatment discontinuation (Chapter 3, Section 3.4.2, page 68), prescriber preference for the prescribed drug (Chapter 5, Section 5.4.2, page 141) and investigator-related factors (methods, specifically ascertainment of drug discontinuation, Chapter 4, Section 4.4.2, page 96). These factors have differential effects between groups of patients treated with different drugs. Clinical and policy implications Although comparative persistence is not a valid measure of the relative benefit-harm balance using currently available TNFα antagonists data, data on persistence and comparative persistence with the TNFα antagonists is useful in interpreting results of other studies, such as effectiveness studies (as described in Section 6.6, page 177). Persistence on the drug, for whatever reason, is critical to determining cost associated with treatment and therefore information on persistence and comparative persistence with the TNFα antagonists is useful for cost-effectiveness analysis and budget planning. Research implications The analysis herein as to whether or not persistence and comparative persistence are valid measures of benefit-harm balance is indirect. Definitive evidence is required from a study that directly tests for the correlation between persistence and effectiveness/adverse effects (using longitudinal data analysis). While all previous studies examined the effect of drug discontinuation on therapeutic benefit, testing for the influence of ineffectiveness on the likelihood to discontinue TNFα antagonists has been neglected. Patients should be followed to ascertain the benefit and harm they experience from the treatment, and determine the effect of lack of benefit or perceived harm on the likelihood of discontinuation.    177 This type of research is challenging and requires access to longitudinal data on therapeutic benefit, harmful events and TNFα antagonist administration dates, as well as techniques for complicated analysis. 6.6 Relevance to Studies of Therapeutic Outcomes  Accurate measurement of drug exposure is essential in pharmacoepidemiologic studies of therapeutic outcomes. Especially important are accurate descriptions of the exposed and unexposed populations, the timing of exposure, and the duration and intensity of exposure. We show that even when drugs within the same therapeutic group are compared, the relative duration of exposure to TNFα antagonists (comparative persistence) is sensitive to methodological issues and therefore the approach used may influence the estimated association between the drug and the study outcome. Sensitivity analysis using different methods to ascertain drug discontinuation is also essential in studies of therapeutic outcomes. Studies of therapeutic outcomes seldom cluster patients by care-providers. It is intuitive that patients treated by the same physician are more similar than patients treated by other physicians, both because of the behaviour of the physician, but also as a result of the patient’s reasons to approach a particular physician. The significance of clustering by care-providing physicians was demonstrated in drug initiating [274,275] and in drug discontinuation ([48], Chapter 3, Section 3.4.2, page 68). Therefore, it seems probable that the prescribing behaviour of physicians varies and therefore influences the length of exposure to the drug. Clustering by care-providing physician can also influence some of the outcomes by affecting time of    178 diagnosis or coding. We assume that ignoring clustering of patients by physician may bias results of therapeutic outcome studies, due to differences in drug persistence (exposure), but also awareness or sensitivity to therapeutic outcomes.     179 BIBLIOGRAPHY 1 Porta M. A dictionary of epidemiology. 5th ed. Oxford, England: Oxford Reference Online, Oxford University Press 2008. 2 Anderson DM, Keith J, Novak PD, et al. Mosby's medical, nursing, & allied health dictionary. 6th ed. Toronto , Ontario , Canada: Elsevier Canada 2002. 3 Cammack R, Attwood TK, Campbell PN, et al. Oxford dictionary of biochemistry and molecular biology. Rev ed. Oxford, England: Oxford Reference Online, Oxford University Press 2006. 4 Jerome D, McKendry RJR. Rheumatology. In: Lock S, Last JM, Dunea G, eds. The Oxford companion to medicine. 3rd ed. Oxford, England: Oxford Reference Online, Oxford University Press; 2001. 5 Xavier JB, Sander C. Principle of system balance for drug interactions. N Engl J Med 2010;362(14):1339-1340. 6 Agency for Healthcare Research and Quality (AHRQ). Glossary of terms. 2012 Available at: http://www.effectivehealthcare.ahrq.gov/index.cfm/glossary-of-terms. Accessed February, 2012. 7 Cramer JA, Roy A, Burrell A, et al. Medication compliance and persistence: Terminology and definitions. Value Health 2008;11(1):44-47. 8 Salas M, Hofman A, Stricker BHC. Confounding by indication: An example of variation in the use of epidemiologic terminology. Am J Epidemiol 1999;149(11):981-983. 9 Grobbee DE, Hoes AW. Confounding and indication for treatment in evaluation of drug treatment for hypertension. BMJ 1997;315(7116):1151-1154. 10 Wolfe F, Flowers N, Burke TA, et al. Increase in lifetime adverse drug reactions, service utilization, and disease severity among patients who will start COX-2 specific inhibitors: Quantitative assessment of channeling bias and confounding by indication in 6689 patients with rheumatoid arthritis and osteoarthritis. J Rheumatol 2002;29(5):1015-1022. 11 Signorello LB, McLaughlin JK, Lipworth L, et al. Confounding by indication in epidemiologic studies of commonly used analgesics. Am J Ther 2002;9(3):199-205. 12 Walker AM. Confounding by indication. Epidemiology 1996;7(4):335-336. 13 Daintith J, Martin EA. A dictionary of science. 6th ed. Oxford, England: Oxford Reference Online, Oxford University Press 2010. 14 Medicare GenerationRx. Glossary of Terms. 2012 Available at: https://www.medicaregenerationrx.com/mp/public/Glossary.jsp. Accessed March, 2012. 15 Kirch W, ed. Encyclopedia of public health. New York, New York, U.S.: Springer 2008. 16 Gerth WC. Compliance and persistence with newer antihypertensive agents. Curr Hypertens Rep 2002;4(6):424-433.    180 17 Andrade SE, Kahler KH, Frech F, et al. Methods for evaluation of medication adherence and persistence using automated databases. Pharmacoepidemiol Drug Saf 2006;15(8):565- 74; discussion 575-7. 18 Hall W. Compliance to Medicines Working Group (CMWG) report to Pharmaceutical Benefits Advisory Committee (PBAC). 2010 Available at: http://www.pbs.gov.au/pbs/pdf- viewer?pdf=%2Fpublication%2Ffactsheets%2Fshared%2F2010-09-20-CMWG-report-to- PBAC.pdf. Accessed June, 2011. 19 The Cochrane Collaboration. Glossary. 2012 Available at: http://www.cochrane.org/glossary/5#letterh. Accessed March, 2012. 20 Hawthorn J, Martin EA. An A-Z of medicinal drugs. 2nd ed. Oxford, England: Oxford Reference Online, Oxford University Press 2010. 21 Kruk ME, Schwalbe N. The relation between intermittent dosing and adherence: preliminary insights. Clin Ther 2006;28(12):1989-1995. 22 Everitt B. The Cambridge dictionary of statistics. 2nd ed. Cambridge, England: Cambridge University Press 2002. 23 Rothman K, Greenland S, Lash T. Modern epidemiology. 3rd ed. Philadelphia, Pennsylvania, U.S.: Wolters Kluwer Health/Lippincott Williams & Wilkins 2008. 24 Martin EA, McFerran TA. A dictionary of nursing. 6th ed. Oxford: Oxford Reference Online, Oxford University Press 2008. 25 British Columbia Ministry of Health. PharmaCare and You. 2011 Available at: http://www.health.gov.bc.ca/pharmacare/generalinfo/generalinfoindex.html. Accessed March, 2012. 26 Peters M. PharmaNet stewardship committee PharmaNet data disclosure [report no 5055-P]. 2007 (Updated 2009);Available at: http://library.bcpharmacists.org/G-PharmaNet/G- 4_PharmaNet_Data_Disclosure/5055-PNet_Data_Disclosure_Requests.pdf. Accessed June, 2011. 27 British Columbia Ministry of Health. General PharmaCare coverage policies. 2011 Available at: http://www.health.gov.bc.ca/pharmacare/policy.html#10. Accessed February, 2012. 28 WHO Collaborating Centre for Drug Statistics Methodology and Norwegian Institute of Public Health. ATC/DDD Index 2012. 2011 Available at: http://www.whocc.no/atc_ddd_index. Accessed January, 2012. 29 American Society of Health-System Pharmacists. AHFS Pharmacologic-Therapeutic Classification System. 2012 Available at: http://www.ahfsdruginformation.com/class/index.aspx. Accessed Febraury, 2012. 30 Halpern MT, Khan ZM, Schmier JK, et al. Recommendations for evaluating compliance and persistence with hypertension therapy using retrospective data. Hypertension 2006;47(6):1039-1048. 31 Myasoedova E, Davis JM,3rd, Crowson CS, et al. Epidemiology of rheumatoid arthritis: Rheumatoid arthritis and mortality. Curr Rheumatol Rep 2010;12(5):379-385.    181 32 Pincus T, Callahan LF. What is the natural history of rheumatoid arthritis? Rheum Dis Clin North Am 1993;19(1):123-151. 33 Wolfe F. The natural history of rheumatoid arthritis. J Rheumatol 1996;44:13-22. 34 Mikuls TR. Co-morbidity in rheumatoid arthritis. Best Prac Res Clin Rheumatol 2003;17(5):729-752. 35 Verstappen SM, Bijlsma JW, Verkleij H, et al. Overview of work disability in rheumatoid arthritis patients as observed in cross-sectional and longitudinal surveys. Arthritis Rheum 2004;51(3):488-497. 36 Burton W, Morrison A, Maclean R, et al. Systematic review of studies of productivity loss due to rheumatoid arthritis. Occup Med 2006;56(1):18-27. 37 Cooper NJ. Economic burden of rheumatoid arthritis: A systematic review. Rheumatology (Oxford) 2000;39(1):28-33. 38 Lundkvist J, Kastang F, Kobelt G. The burden of rheumatoid arthritis and access to treatment: Health burden and costs. Eur J Health Econ 2008;8(Suppl 2):S49-S60. 39 Anderson PJ. Tumor necrosis factor inhibitors: Clinical implications of their different immunogenicity profiles. Semin Arthritis Rheum 2005;34(5 Suppl 1):19-22. 40 Arora T, Padaki R, Liu L, et al. Differences in binding and effector functions between classes of TNF antagonists. Cytokine 2009;45(2):124-131. 41 Dost T. The pharmacokinetics of biologic agents. Antiinflamm Antiallergy Agents Med Chem 2010;9(1):3-7. 42 Nestorov I. Clinical pharmacokinetics of TNF antagonists: How do they differ? Semin Arthritis Rheum 2005;34(5 Suppl 1):12-18. 43 Rigby WF. Drug insight: Different mechanisms of action of tumor necrosis factor antagonists-passive-aggressive behavior? Nat Clin Pract Rheumatol 2007;3(4):227-233. 44 Feldmann M, Maini RN. Anti-TNF alpha therapy of rheumatoid arthritis: what have we learned? Annu Rev Immunol 2001;19:163-196. 45 Wolfe F. The epidemiology of drug treatment failure in rheumatoid arthritis. Baillieres Clin Rheumatol 1995;9(4):619-632. 46 Janhsen K, Vlahović-Palčevski V, Elseviers M, et al. Cross national comparison of drug utilization activities: Project description; survey protocol. Available at: http://www.pharmacoepi.org/eurodurg/workgr/cross_national/ProjectPlanCNC final.doc. Accessed July, 2012. 47 Garvie PA, Flynn PM, Belzer M, et al. Psychological factors, beliefs about medication, and adherence of youth with human immunodeficiency virus in a multisite directly observed therapy pilot study. J Adolesc Health 2011;48(6):637-640. 48 Zhang J, Shan Y, Reed G, et al. Thresholds in disease activity for switching biologics in rheumatoid arthritis patients: Experience from a large US cohort. Arthritis Care Res 2011;63(12):1672-1679.    182 49 Blum MA, Koo D, Doshi JA. Measurement and rates of persistence with and adherence to biologics for rheumatoid arthritis: A systematic review. Clin Ther 2011;33(7):901-913. 50 Koncz T, Pentek M, Brodszky V, et al. Adherence to biologic DMARD therapies in rheumatoid arthritis. Expert Opin Biol Ther 2010;10(9):1367-1378. 51 DiMatteo MR, Giordani PJ, Lepper HS, et al. Patient adherence and medical treatment outcomes: A meta-analysis. Med Care 2002;40(9):794-811. 52 Cramer JA, Benedict A, Muszbek N, et al. The significance of compliance and persistence in the treatment of diabetes, hypertension and dyslipidaemia: A review. Int J Clin Pract 2008;62(1):76-87. 53 Simpson SH, Eurich DT, Majumdar SR, et al. A meta-analysis of the association between adherence to drug therapy and mortality. BMJ 2006;333(7557):15. 54 Bae JW, Guyer W, Grimm K, et al. Medication persistence in the treatment of HIV infection: A review of the literature and implications for future clinical care and research. AIDS 2011;25(3):279-290. 55 World Health Organization (WHO). Adherence to long-term therapies: Evidence for action. 2003 Available at: http://www.who.int/chp/knowledge/publications/adherence_full_report.pdf. Accessed June, 2011. 56 Strom LS, Kimmel SE. Textbook of pharmacoepidemiology. 4th ed. West Sussez, England: John Wiley & Sons, Ltd 2007. 57 Peterson AM, Nau DP, Cramer JA, et al. A checklist for medication compliance and persistence studies using retrospective databases. Value Health 2007;10(1):3-12. 58 Leslie S, Gwadry-Sridhar F, Thiebaud P, et al. Calculating medication compliance, adherence, and persistence in administrative pharmacy claims databases. Pharmaceutical Programming 2008;1(1):13-19. 59 Caetano PA, Lam JM, Morgan SG. Toward a standard definition and measurement of persistence with drug therapy: Examples from research on statin and antihypertensive utilization. Clin Ther 2006;28(9):1411-24; discussion 1410. 60 Motulsky H. Chapter 6: Survival curves. In: Motulsky H, ed. Intuitive biostatistics. 1st ed. Oxford, England: Oxford University Press; 1995:53-60. 61 Borah BJ, Huang X, Zarotsky V, et al. Trends in RA patients' adherence to subcutaneous anti-TNF therapies and costs. Curr Med Res Opin 2009;25(6):1365-1377. 62 Yazici Y, Krasnokutsky S, Barnes JP, et al. Changing patterns of tumor necrosis factor inhibitor use in 9074 patients with rheumatoid arthritis. J Rheumatol 2009;36(5):907-913. 63 Curkendall S, Patel V, Gleeson M, et al. Compliance with biologic therapies for rheumatoid arthritis: Do patient out-of-pocket payments matter? Arthritis Rheum 2008;59(10):1519- 1526.    183 64 Grijalva CG, Chung CP, Arbogast PG, et al. Assessment of adherence to and persistence on disease-modifying antirheumatic drugs (DMARDs) in patients with rheumatoid arthritis. Med Care 2007;45(10 Suppl 2):S66-S76. 65 Abarca J, Malone DC, Armstrong EP, et al. Longitudinal analysis of the use of etanercept versus infliximab determined from medical chart audit. JMCP 2004;10(6):538-542. 66 Duclos M, Gossec L, Ruyssen-Witrand A, et al. Retention rates of tumor necrosis factor blockers in daily practice in 770 rheumatic patients. J Rheumatol 2006;33(12):2433-2438. 67 Garcia Rodriguez LA, Perez Gutthann S. Use of the UK General Practice Research Database for pharmacoepidemiology. Br J Clin Pharmacol 1998;45(5):419-425. 68 Bertoli AM, Strusberg I, Baravalle M, et al. Rate and causes of infliximab discontinuation in patients with rheumatoid arthritis in a private clinical practice. J Clin Rheumatol 2008;14(6):313-317. 69 Ducoulombier V, Solau E, Coquerelle P, et al. Long-term results of infliximab therapy in rheumatoid arthritis: Experience acquired by the North-Pas-de-Calais hospital network. Joint Bone Spine 2007;74(1):56-59. 70 Ray WA. Evaluating medication effects outside of clinical trials: New-user designs. Am J Epidemiol 2003;158(9):915-920. 71 Schmeichel-Mueller C, Bartels A, Silver J, et al. Frequency of select anti-TNF administration or re-fill in patients with rheumatoid arthritis [abstract]. Value Health 2011;14(3):A135. 72 Wu E, Chen L, Birnbaum H, et al. Retrospective claims data analysis of dosage adjustment patterns of TNF antagonists among patients with rheumatoid arthritis. Curr Med Res Opin 2008;24(8):2229-2240. 73 Nielsen LH, Keiding N. Validation of methods for identifying discontinuation of treatment from prescription data. J Roy Statist Soc Ser C 2010;59(4):707-722. 74 Cooke CE, Lee HY, Tong YP, et al. Persistence with injectable antidiabetic agents in members with type 2 diabetes in a commercial managed care organization. Curr Med Res Opin 2010;26(1):231-238. 75 Maini RN, Breedveld FC, Kalden JR, et al. Therapeutic efficacy of multiple intravenous infusions of anti-tumor necrosis factor alpha monoclonal antibody combined with low- dose weekly methotrexate in rheumatoid arthritis. Arthritis Rheum 1998;41(9):1552-1563. 76 den Broeder A, van de Putte L, Rau R, et al. A single dose, placebo controlled study of the fully human anti-tumor necrosis factor-alpha antibody adalimumab (D2E7) in patients with rheumatoid arthritis. J Rheumatol 2002;29(11):2288-2298. 77 Sikka R, Xia F, Aubert RE. Estimating medication persistency using administrative claims data. Am J Manag Care 2005;11(7):449-457. 78 Mahic M, Skurtveit S, Selmer R, et al. Prevalence, incidence and persistence of etanercept and adalimumab in Norway 2005-2009. Pharmacoepidemiol Drug Saf 2011;20(5):457- 463.    184 79 Agarwal SK, Glass RJ, Shadick NA, et al. Predictors of discontinuation of tumor necrosis factor inhibitors in patients with rheumatoid arthritis. J Rheumatol 2008;35(9):1737-1744. 80 Aalen OO. A model for nonparametric regression analysis of counting processes. Mathematical statistics and probability theory. Lecture Notes in Statistics 1980;(2):1-25. 81 Aalen OO. A linear regression model for the analysis of life times. Stat Med 1989;8(8):907- 925. 82 Cao H. A Comparison between the additive and multiplicative risk models [Dissertation]. Quebec, Canada: l’Universit´e Laval 2005. 83 Marton C, Moravan V. BC Health Data Warehouse (BC HDW) from effectively  searching for health statistics [Conference proceeding]. 2006 Available at: http://www.viha.ca/NR/rdonlyres/EA187CF8-C0C0-4699-9097- 6E5BF7DF1665/0/bchealthdatawarehouse.pdf. Accessed September, 2011. 84 Hu W. Summary report [unpublished draft] Victoria,  British Columbia, Canada: Program Monitoring and Information Management Branch, Resource Management Division British Columbia Ministry of Health and Ministry Responsible for Seniors. 1996. 85 The Canadian Institute for Health Information (CIHI). Discharge Abstract Database (DAD) re-abstraction studies. 2010 Available at: https://secure.cihi.ca/estore/productSeries.htm?pc=PCC228. Accessed September, 2011. 86 Garrison LP,Jr, Neumann PJ, Erickson P, et al. Using real-world data for coverage and payment decisions: the ISPOR Real-World Data Task Force report. Value Health 2007;10(5):326-335. 87 Motheral BR, Fairman KA. The use of claims databases for outcomes research: Rationale, challenges, and strategies. Clin Ther 1997;19(2):346-366. 88 Paul JE, Weis KA, Epstein RA. Data bases for variations research. Med Care 1993;31(5 Suppl):YS96-YS102. 89 Suissa S, Garbe E. Primer: Administrative health databases in observational studies of drug effects – Advantages and disadvantages. Nat Clin Pract Rheumatol 2007;3(12):725-732. 90 Woolf SH, Rothemich SF, Johnson RE, et al. Selection bias from requiring patients to give consent to examine data for health services research. Arch Fam Med 2000;9(10):1111- 1118. 91 Adamis D, Martin FC, Treloar A, et al. Capacity, consent, and selection bias in a study of delirium. J Med Ethics 2005;31(3):137-143. 92 Ross LL, Ross NP, Fisher ES, Bubolz TA. Chapter 5: Strengths and weaknesses of health insurance data systems for assessing outcomes. In: Gelijns A, ed. Modern methods of clinical investigation (Medical innovation at the crossroads, volume 1) Washington, D.C., US: National Academy Press; 1990:47. 93 Sax MJ. Essential steps and practical applications for database studies. J Manag Care Pharm 2005;11(1 Suppl A):S5-S8.    185 94 Kramer MS, Shapiro SH. Scientific challenges in the application of randomized trials. JAMA 1984;252(19):2739-2745. 95 McCarney R, Warner J, Iliffe S, et al. The Hawthorne effect: A randomised, controlled trial. BMC Med Res Methodol 2007;7:30. 96 Katz JN, Barrett J, Liang MH, et al. Sensitivity and positive predictive value of Medicare Part B physician claims for rheumatologic diagnoses and procedures. Arthritis Rheum 1997;40(9):1594-1600. 97 Losina E, Barrett J, Baron JA, et al. Accuracy of Medicare claims data for rheumatologic diagnoses in total hip replacement recipients. J Clin Epidemiol 2003;56(6):515-519. 98 Singh JA, Holmgren AR, Noorbaloochi S. Accuracy of Veterans Administration databases for a diagnosis of rheumatoid arthritis. Arthritis Rheum 2004;51(6):952-957. 99 Lix L, Yogendran M, Burchill C, et al. Defining and validating chronic diseases: An administrative  data approach. 2006 Available at: http://mchp- appserv.cpe.umanitoba.ca/reference/chronic.disease.pdf. Accessed September, 2009. 100 Statistics Canada. Health Statistics Division. Canadian community health survey, Cycle 1.1. 2000 Available at: http://data.library.ubc.ca/java/jsp/database/production/detail.jsp?id=776. Accessed September, 2011. 101 Lacaille D, Anis AH, Guh DP, et al. Gaps in care for rheumatoid arthritis: A population study. Arthritis Rheum 2005;53(2):241-248. 102 Lacaille D, Guh DP, Abrahamowicz M, et al. Use of nonbiologic disease-modifying antirheumatic drugs and risk of infection in patients with rheumatoid arthritis. Arthritis Rheum 2008;59(8):1074-1081. 103 Li X, Guh D, Lacaille D, et al. The impact of cost sharing of prescription drug expenditures on health care utilization by the elderly: Own- and cross-price elasticities. Health Policy 2007;82(3):340-347. 104 Tang B, Rahman M, Waters HC, et al. Treatment persistence with adalimumab, etanercept, or infliximab in combination with methotrexate and the effects on health care costs in patients with rheumatoid arthritis. Clin Ther 2008;30(7):1375-1384. 105 Braun-Moscovici Y, Markovits D, Rozin A, et al. Anti-tumor necrosis factor therapy: 6 year experience of a single center in northern Israel and possible impact of health policy on results. Israel med Assoc J 2008;10(4):277-281. 106 Brocq O, Roux CH, Albert C, et al. TNFalpha antagonist continuation rates in 442 patients with inflammatory joint disease. Joint Bone Spine 2007;74(2):148-154. 107 Carter C, Jang B, Buysman E, et al. Two-year longitudinal study of biologic switching and discontinuation among rheumatoid arthritis patients in a health plan [abstract]. J Manag Care Pharm 2010;16(2):165-166. 108 Flendrie M, Creemers MCW, Welsing PMJ, et al. Survival during treatment with tumour necrosis factor blocking agents in rheumatoid arthritis. Ann Rheum Dis 2003;62(2):i30- i33.    186 109 Hetland ML, Lindegaard HM, Hansen A, et al. Do changes in prescription practice in patients with rheumatoid arthritis treated with biological agents affect treatment response and adherence to therapy? Results from the nationwide Danish DANBIO Registry. Ann Rheum Dis 2008;67(7):1023-1026. 110 Hetland ML, Christensen IJ, Tarp U, et al. Direct comparison of treatment responses, remission rates, and drug adherence in patients with rheumatoid arthritis treated with adalimumab, etanercept, or infliximab: Results from eight years of surveillance of clinical practice in the nationwide Danish DANBIO registry. Arthritis Rheum 2010;62(1):22-32. 111 Blom M, Kievit W, Fransen J, et al. The reason for discontinuation of the first tumor necrosis factor (TNF) blocking agent does not influence the effect of a second TNF blocking agent in patients with rheumatoid arthritis. J Rheumatol 2009;36(10):2171-2177. 112 Conti F, Scrivo R, Spinelli FR, et al. Outcome in patients with rheumatoid arthritis switching TNF-alpha antagonists: A single center, observational study over an 8-year period. Clin Exp Rheumatol 2009;27(3):540-541. 113 Finckh A, Simard JF, Gabay C, et al. Evidence for differential acquired drug resistance to anti-tumour necrosis factor agents in rheumatoid arthritis. Ann Rheum Dis 2006;65(6):746- 752. 114 Finckh A, Dehler S, Gabay C. The effectiveness of leflunomide as a co-therapy of tumour necrosis factor inhibitors in rheumatoid arthritis: A population-based study. Ann Rheum Dis 2009;68(1):33-39. 115 Heiberg MS, Koldingsnes W, Mikkelsen K, et al. The comparative one-year performance of anti-tumor necrosis factor alpha drugs in patients with rheumatoid arthritis, psoriatic arthritis, and ankylosing spondylitis: Results from a longitudinal, observational, multicenter study. Arthritis Rheum 2008;59(2):234-240. 116 Genevay S, Finckh A, Ciurea A, et al. Tolerance and effectiveness of anti-tumor necrosis factor alpha therapies in elderly patients with rheumatoid arthritis: A population-based cohort study. Arthritis Care Res 2007;57(4):679-685. 117 Gomez-Reino JJ, Carmona L, the BIOBADASER Group. Switching TNF antagonists in patients with chronic arthritis: An observational study of 488 patients over a four-year period. Arthritis Res Ther 2006;8(1):R29. 118 Vegter S, Nguyen NH, Visser ST, et al. Compliance, persistence, and switching patterns for ACE inhibitors and ARBs. Am J Manag Care 2011;17(9):609-616. 119 M B, X R, Prous J.R. Gateways to clinical trials. Methods Find Exp Clin Pharmacol 2007;29(2):153-173. 120 Bendtzen K, Geborek P, Svenson M, et al. Individualized monitoring of drug bioavailability and immunogenicity in rheumatoid arthritis patients treated with the tumor necrosis factor alpha inhibitor infliximab. Arthritis Rheum 2006;54(12):3782-3789. 121 Ducourau E, Mulleman D, Paintaud G, et al. Antibodies toward infliximab are associated with low infliximab concentration at treatment initiation and poor infliximab maintenance in rheumatic diseases. Arthritis Res Ther 2011;13(3):R105.    187 122 Hyrich KL, Symmons DP, Watson KD, et al. Comparison of the response to infliximab or etanercept monotherapy with the response to cotherapy with methotrexate or another disease-modifying antirheumatic drug in patients with rheumatoid arthritis: Results from the British Society for Rheumatology Biologics Register. Arthritis Rheum 2006;54(6):1786-1794. 123 Klareskog L, van der Heijde D, de Jager JP, et al. Therapeutic effect of the combination of etanercept and methotrexate compared with each treatment alone in patients with rheumatoid arthritis: Double-blind randomised controlled trial. Lancet 2004;363(9410):675-681. 124 Breedveld FC, Weisman MH, Kavanaugh AF, et al. The PREMIER study: A multicenter, randomized, double-blind clinical trial of combination therapy with adalimumab plus methotrexate versus methotrexate alone or adalimumab alone in patients with early, aggressive rheumatoid arthritis who had not had previous methotrexate treatment. Arthritis Rheum 2006;54(1):26-37. 125 Bosco JL, Silliman RA, Thwin SS, et al. A most stubborn bias: No adjustment method fully resolves confounding by indication in observational studies. J Clin Epidemiol 2010;63(1):64-74. 126 Glynn RJ, Schneeweiss S, Sturmer T. Indications for propensity scores and review of their use in pharmacoepidemiology. Basic Clin Pharmacol Toxicol 2006;98(3):253-259. 127 Rubin DB. Estimating causal effects from large data sets using propensity scores. Ann Intern Med 1997;127(8 Part 2):757-763. 128 Sturmer T, Schneeweiss S, Brookhart MA, et al. Analytic strategies to adjust confounding using exposure propensity scores and disease risk scores: Nonsteroidal antiinflammatory drugs and short-term mortality in the elderly. Am J Epidemiol 2005;161(9):891-898. 129 Schneeweiss S, Rassen JA, Glynn RJ, et al. High-dimensional propensity score adjustment in studies of treatment effects using health care claims data. Epidemiology 2009;20(4):512-522. 130 Johnston KM, Gustafson P, Levy AR, et al. Use of instrumental variables in the analysis of generalized linear models in the presence of unmeasured confounding with applications to epidemiological research. Stat Med 2008;27(9):1539-1556. 131 Brookhart MA, Schneeweiss S. Preference-based instrumental variable methods  for the estimation of treatment effects: Assessing validity and interpreting results. Int J Biostat 2007;3(1):14. 132 Schneeweiss S, Solomon DH, Wang PS, et al. Simultaneous assessment of short-term gastrointestinal benefits and cardiovascular risks of selective cyclooxygenase 2 inhibitors and nonselective nonsteroidal antiinflammatory drugs: An instrumental variable analysis. Arthritis Rheum 2006;54(11):3390-3398. 133 Schneeweiss S, Avorn J. A review of uses of health care utilization databases for epidemiologic research on therapeutics. J Clin Epidemiol 2005;58(4):323-337.    188 134 Erkens JA, Panneman MM, Klungel OH, et al. Differences in antihypertensive drug persistence associated with drug class and gender: A PHARMO study. Pharmacoepidemiol Drug Saf 2005;14(11):795-803. 135 Degli Esposti E, Sturani A, Di Martino M, et al. Long-term persistence with antihypertensive drugs in new patients. J Hum Hypertens 2002;16(6):439-444. 136 Abughosh SM, Kogut SJ, Andrade SE, et al. Persistence with lipid-lowering therapy: Influence of the type of lipid-lowering agent and drug benefit plan option in elderly patients. J Manag Care Pharm 2004;10(5):404-411. 137 Iikuni N, Sato E, Hoshi M, et al. The influence of sex on patients with rheumatoid arthritis in a large observational cohort. J Rheumatol 2009;36(3):508-511. 138 Owen CG, Carey IM, de Wilde S, et al. Persistency with medical treatment for glaucoma and ocular hypertension in the United Kingdom: 1994-2005. Eye 2009;23(5):1098-1110. 139 Arias A, Schargel K, Ussa F, et al. Patient persistence with first-line antiglaucomatous monotherapy. Clin Ophthalmol 2010;4:261-267. 140 Chilton F, Collett RA. Treatment choices, preferences and decision-making by patients with rheumatoid arthritis. Muscoskel Care 2008;6(1):1-14. 141 British Columbia Ministry of Health. Your level of Fair PharmaCare coverage in Fair PharmaCare. 2011 Available at: http://www.health.gov.bc.ca/pharmacare/plani/planiindex.html#7. Accessed March, 2011. 142 Garcia-Gonzalez A, Richardson M, Garcia Popa-Lisseanu M, et al. Treatment adherence in patients with rheumatoid arthritis and systemic lupus erythematosus. Clin Rheumatol 2008;27(7):883-889. 143 Lorish CD, Richards B, Brown S. Missed medication doses in rheumatic arthritis patients: Intentional and unintentional reasons. Arthritis Care Res 1989;2(1):3-9. 144 Foster SA, Foley KA, Meadows ES, et al. Adherence and persistence with teriparatide among patients with commercial, Medicare, and Medicaid insurance. Osteoporos Int 2011;22(2):551-557. 145 Ellis JJ, Erickson SR, Stevenson JG, et al. Suboptimal statin adherence and discontinuation in primary and secondary prevention populations. J Gen Intern Med 2004;19(6):638-645. 146 Friedman O, McAlister FA, Yun L, et al. Antihypertensive drug persistence and compliance among newly treated elderly hypertensives in Ontario. Am J Med 2010;123(2):173-181. 147 Yeaw J, Benner JS, Walt JG, et al. Comparing adherence and persistence across 6 chronic medication classes. J Manag Care Pharm 2009;15(9):728-740. 148 Kimmick G, Anderson R, Camacho F, et al. Adjuvant hormonal therapy use among insured, low-income women with breast cancer. J Clin Oncol 2009;27(21):3445-3451.    189 149 Staples MP, March L, Lassere M, et al. Health-related quality of life and continuation rate on first-line anti-tumour necrosis factor therapy among rheumatoid arthritis patients from the Australian Rheumatology Association Database. Rheumatology (Oxford) 2011;50(1):166-175. 150 Kristensen LE, Saxne T, Nilsson JA, et al. Impact of concomitant DMARD therapy on adherence to treatment with etanercept and infliximab in rheumatoid arthritis. Results from a six-year observational study in southern Sweden. Arthritis Res Ther 2006;8(6):R174. 151 Marchesoni A, Zaccara E, Gorla R, et al. TNF-alpha antagonist survival rate in a cohort of rheumatoid arthritis patients observed under conditions of standard clinical practice. Ann N Y Acad Sci 2009;1173(1):837-846. 152 Mattey DL, Dawes PT, Hassell AB, et al. Effect of psychological distress on continuation of anti-tumor necrosis factor therapy in patients with rheumatoid arthritis. J Rheumatol 2010;37(10):2021-2024. 153 Faraawi R, Hack F, Bensen W, et al. Prior use of leflunomide increases the hazard ratio of discontinuation  of infliximab in a Canadian open label prospective study with overall  drop-out rates similar to other biologic therapies for the treatment of  rheumatoid arthritis [abstract]. J Rheumatol 2009;36(11):2571-2572. 154 Soliman MM, Ashcroft DM, Watson KD, et al. Benefit of concomitant dmard use on the persistence with anti-TNF therapies in rheumatoid arthritis: Results from the British Society of  Rheumatology Biologics Register (BSRBR) [abstract]. Rheumatology (Oxford) 2010;49(Suppl 1):3-4. 155 Voulgari PV, Alamanos Y, Nikas SN, et al. Infliximab therapy in established rheumatoid arthritis: An observational study. Am J Med 2005;118(5):515-520. 156 Kievit W, Fransen J, Adang EMM, et al. Long-term effectiveness and safety of TNF- blocking agents in daily clinical practice: Results from the Dutch Rheumatoid Arthritis Monitoring Register. Rheumatology (Oxford) 2011;50(1):196-203. 157 Bensen B, Atkins C, Sheriff M, et al. Profile of patients with rheumatoid arthritis treated with infliximab in Canada-trends toward less DMARD use prior to a biologic, earlier use of infliximab and differences in baseline disease [abstract]. J Rheumatol 2011;38(6):1155. 158 Buchanan J, Rahman MU, Doyle MK, et al. Historical trends of patient characteristics in anti-TNF clinical trials for rheumatoid arthritis: An analysis of the literature over the past 16 years [abstract]. Arthritis Rheum 2009;60(Supple 10):1012. 159 Hyrich KL, Watson KD, Lunt M, et al. Changes in disease characteristics and response rates among patients in the United Kingdom starting anti-tumour necrosis factor therapy for rheumatoid arthritis between 2001 and 2008. Rheumatology (Oxford) 2011;50(1):117- 123. 160 Lee SJ, Chang H, Yazici Y, et al. Utilization trends of tumor necrosis factor inhibitors among patients with rheumatoid arthritis in a United States observational cohort study. J Rheumatol 2009;36(8):1611-1617.    190 161 Fernandez-Nebro A, Irigoyen MV, Urena I, et al. Effectiveness, predictive response factors, and safety of anti-tumor necrosis factor (TNF) therapies in anti-TNF-naive rheumatoid arthritis. J Rheumatol 2007;34(12):2334-2342. 162 Amuah JE, Hogan DB, Eliasziw M, et al. Persistence with cholinesterase inhibitor therapy in a population-based cohort of patients with Alzheimer's disease. Pharmacoepidemiol Drug Saf 2010;19(7):670-679. 163 Taira DA, Gelber RP, Davis J, et al. Antihypertensive adherence and drug class among Asian Pacific Americans. Ethn Health 2007;12(3):265-281. 164 Jones TJ, Petrella RJ, Crilly R. Determinants of persistence with weekly bisphosphonates in patients with osteoporosis. J Rheumatol 2008;35(9):1865-1873. 165 Hansen DG, Vach W, Rosholm JU, et al. Early discontinuation of antidepressants in general practice: Association with patient and prescriber characteristics. Fam Pract 2004;21(6):623-629. 166 Catalan VS, Couture JA, LeLorier J. Predictors of persistence of use of the novel antidiabetic agent acarbose. Arch Intern Med 2001;161(8):1106-1112. 167 Wilensky J, Fiscella RG, Carlson AM, et al. Measurement of persistence and adherence to regimens of IOP-lowering glaucoma medications using pharmacy claims data. Am J Ophthalmol 2006;141(1 Suppl):S28-S33. 168 Grymonpre R, Cheang M, Fraser M, et al. Validity of a prescription claims database to estimate medication adherence in older persons. Med Care 2006;44(5):471-477. 169 British Columbia Ministry of Health. Section 10.4: Maximum days’ supply policy, in  Policies and procedures manual (Pharamcare and PharmaNet). 2003 Available at: http://www.health.gov.bc.ca/pharmacare/generalinfo/policy/. Accessed September, 2011. 170 Du Pan SM, Dehler S, Ciurea A, et al. Comparison of drug retention rates and causes of drug discontinuation between anti-tumor necrosis factor agents in rheumatoid arthritis. Arthritis Rheum 2009;61(5):560-568. 171 Pocock JM, Albreshni S, Slack R, et al. A comparison of drug survival times for infliximab, adalimumab and etanercept [abstract]. Rheumatology (Oxford) 2009;48(Suppl 1):69. 172 Filippini M, Bazzani C, Favalli EG, et al. Efficacy and safety of anti-tumour necrosis factor in elderly patients with rheumatoid arthritis: An observational study. Clin Rev Allergy Immunol 2010;38(2-3):90-96. 173 Smith T, Smith B. Kaplan Meier and Cox proportional hazards modeling: Hands on survival analysis. 2004 Available at: http://www.lexjansen.com/wuss/2004/hands_on_workshops/i_how_kaplan_meier_and_co x_p.pdf. Accessed April, 2012. 174 Brennan FM, Maini RN, Feldmann M. TNF alpha – A pivotal role in rheumatoid arthritis? Br J Rheumatol 1992;31(5):293-298. 175 Quan H, Sundararajan V, Halfon P, et al. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med Care 2005;43(11):1130-1139.    191 176 Charlson ME, Pompei P, Ales KL, et al. A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. J Chronic Dis 1987;40(5):373-383. 177 Hosmer DW, Lemeshow S. Applied survival analysis: Regression modeling of time to event data. New York, New York, U.S.: Wiley 1999:386. 178 SAS Institute Inc. SAS (Statistical Analysis System) software package for Windows software package. Copyright © 2002-2003;9.1.3. 179 Ornbjerg LM, Ostergaard M, Boyesen P, et al. Discontinuation of TNF-inhibitor treatment in clinical practice has a negative impact on radiographic progression 2 years after initiation of therapy. Results from the nationwide Danish DANBIO registry [abstract]. Arthritis Rheum 2010;62(Suppl 10):1444. 180 Wendling D, Materne G-, Michel F, et al. Infliximab continuation rates in patients with rheumatoid arthritis in everyday practice. Joint Bone Spine 2005;72(4):309-312. 181 Vander Cruyssen B, Van Looy S, Wyns B, et al. Four-year follow-up of infliximab therapy in rheumatoid arthritis patients with long-standing refractory disease: Attrition and long-term evolution of disease activity. Arthritis Res Ther 2006;8(4):R112. 182 Jarry JL, Coambs RB, De Maio FG, et al. Projected patterns of compliance associated with Remicade (infliximab) and Enbrel (etanercept) (AB23) [abstract]. J Rheumatol 2002;29(7):1564. 183 Jarry JL, Coambs RB, De Maio FG, et al. Expected compliance profiles of the new antirheumatics, infliximab and etanercept [poster]. 2001 Available at: http://www.healthpromoresearch.com/images/ILAR%20poster%20for%20print.pdf. Accessed December, 2010. 184 Chapman RH, Petrilla AA, Benner JS, et al. Predictors of adherence to concomitant antihypertensive and lipid-lowering medications in older adults: A retrospective, cohort study. Drugs Aging 2008;25(10):885-892. 185 Brookhart MA, Patrick AR, Schneeweiss S, et al. Physician follow-up and provider continuity are associated with long-term medication adherence: A study of the dynamics of statin use. Arch Intern Med 2007;167(8):847-852. 186 Shukla R, Vender RB. Pharmacology of TNF inhibitors. In: Weinberg JM, Buchholz R, eds. TNF-alpha inhibitors (Milestones in drug therapy) Basel, Switzerland: Birkhäuser Verlag; 2006:23-44. 187 Emi Aikawa N, de Carvalho JF, Artur Almeida Silva C, et al. Immunogenicity of anti- TNF-alpha agents in autoimmune diseases. Clin Rev Allergy Immunol 2010;38(2-3):82-89. 188 Tracey D, Klareskog L, Sasso EH, et al. Tumor necrosis factor antagonist mechanisms of action: A comprehensive review. Pharmacol Ther 2008;117(2):244-279. 189 Furst DE, Wallis R, Broder M, et al. Tumor necrosis factor antagonists: Different kinetics and/or mechanisms of action may explain differences in the risk for developing granulomatous infection. Semin Arthritis Rheum 2006;36(3):159-167.    192 190 Ware CF. Network communications: Lymphotoxins, LIGHT, and TNF. Annu Rev Immunol 2005;23(1):787-819. 191 Zink A, Listing J, Kary S, et al. Treatment continuation in patients receiving biological agents or conventional DMARD therapy. Ann Rheum Dis 2005;64(9):1274-1279. 192 Miller K, Conner TM, Bunz TJ, et al. A preliminary analysis of biologics medication- taking behavior among rheumatoid arthritis patients [abstract]. Value Health 2009;12(3):A73. 193 Kristensen LE, Saxne T, Geborek P. The LUNDEX, a new index of drug efficacy in clinical practice: Results of a five-year observational study of treatment with infliximab and etanercept among rheumatoid arthritis patients in Southern Sweden. Arthritis Rheum 2006;54(2):600-606. 194 Ostergaard M, Unkerskov J, Linde L, et al. Low remission rates but long drug survival in rheumatoid arthritis patients treated with infliximab or etanercept: Results from the nationwide Danish DANBIO database. Scand J Rheumatol 2007;36(2):151-154. 195 Choo PW, Rand CS, Inui TS, et al. Validation of patient reports, automated pharmacy records, and pill counts with electronic monitoring of adherence to antihypertensive therapy. Med Care 1999;37(9):846-857. 196 Hansen RA, Kim MM, Song L, et al. Comparison of methods to assess medication adherence and classify nonadherence. Ann Pharmacother 2009;43(3):413-422. 197 McKenzie DA, Semradek J, McFarland BH, et al. The validity of medicaid pharmacy claims for estimating drug use among elderly nursing home residents: The Oregon experience. J Clin Epidemiol 2000;53(12):1248-1257. 198 Sedgwick P. Bias in clinical trials. BMJ 2011;343:d4176. 199 Halfon N, Newacheck PW. Trends in the hospitalization for acute childhood asthma, 1970-84. Am J Public Health 1986;76(11):1308-1311. 200 Gergen PJ, Weiss KB. Changing patterns of asthma hospitalization among children: 1979 to 1987. JAMA 1990;264(13):1688-1692. 201 Veen ND, Selten JP, Schols D, et al. Diagnostic stability in a Dutch psychosis incidence cohort. Br J Psychiatry 2004;185:460-464. 202 Vollmer WM, Osborne ML, Buist AS. Uses and limitations of mortality and health care utilization statistics in asthma research. Am J Respir Crit Care Med 1994;149(2 Part 2):S79-S87; discussion S88-S90. 203 Fu JB, Kau TY, Severson RK, et al. Lung cancer in women: Analysis of the national surveillance, epidemiology, and end results database. Chest 2005;127(3):768-777. 204 Todd J, Bertoch D, Dolan S. Use of a large national database for comparative  evaluation of the effect of a bronchiolitis/viral pneumonia clinical care guideline  on patient outcome and resource utilization. Arch Pediatr Adolesc Med 2002;156(11):1086-1090. 205 Dzakpasu S, Joseph KS, Huang L, et al. Decreasing diagnoses of birth asphyxia in Canada: Fact or artifact. Pediatrics 2009;123(4):e668-e672.    193 206 Tedeschi P, Griffith JR. Classification of hospital patients as "surgical." Implications of the shift to ICD-9-CM. Med Care 1984;22(3):189-192. 207 eCPS (online Compendium of Pharmaceuticals and Specialties). Remicade® (infliximab). 2008. Accessed October, 2008. 208 Dezii CM. Persistence with drug therapy: A practical approach using administrative claims data. Manag Care 2001;10(2):42-45. 209 Wolfe F, Michaud K, Stephenson B, et al. Toward a definition and method of assessment of treatment failure and treatment effectiveness: The case of leflunomide versus methotrexate. J Rheumatol 2003;30(8):1725-1732. 210 Kievit W, Fransen J, Oerlemans AJ, et al. The efficacy of anti-TNF in rheumatoid arthritis, a comparison between randomised controlled trials and clinical practice. Ann Rheum Dis 2007;66(11):1473-1478. 211 Sokka T, Pincus T. Eligibility of patients in routine care for major clinical trials of anti- tumor necrosis factor alpha agents in rheumatoid arthritis. Arthritis Rheum 2003;48(2):313-318. 212 Zink A, Strangfeld A, Schneider M, et al. Effectiveness of tumor necrosis factor inhibitors in rheumatoid arthritis in an observational cohort study: Comparison of patients according to their eligibility for major randomized clinical trials. Arthritis Rheum 2006;54(11):3399- 3407. 213 Li P, Blum MA, Von Feldt J, et al. Adherence, discontinuation, and switching of biologic therapies in Medicaid enrollees with rheumatoid arthritis. Value Health 2010;13(6):805- 812. 214 Chang HC, Chen LC, Tseng HL, et al. The long-term utilization and safety of biological agents in treating rheumatoid arthritis patients - A population-based case study in Southern Taiwan [abstract]. Value Health 2010;13(7):A557. 215 Ogale S, Hitraya E, Henk HJ. Patterns of biologic agent utilization among patients with rheumatoid arthritis: A retrospective cohort study. BMC Musculoskelet Disord 2011;12:204. 216 Schmeichel-Mueller C, Buysman E, Bolge S, et al. Definitions of anti-TNF discontinuation may impact understanding of real-world utilization patterns [abstract]. Value Health 2011;14(3):A71. 217 Greevy RA,Jr, Huizinga MM, Roumie CL, et al. Comparisons of persistence and durability among three oral antidiabetic therapies using electronic prescription-fill data: The impact of adherence requirements and stockpiling. Clin Pharmacol Ther 2011;90(6):813-819. 218 Hudson M, Rahme E, Richard H, et al. Comparison of measures of medication persistency using a prescription drug database. Am Heart J 2007;153(1):59-65. 219 Cramer JA, Silverman SL, Gold DT. Methodological considerations in using claims databases to evaluate persistence with bisphosphonates for osteoporosis. Curr Med Res Opin 2007;23(10):2369-2377.    194 220 Mabotuwana T, Warren J. A framework for assessing adherence and persistence to long- term medication. Stud Health Technol Inform 2009;150:547-551. 221 Rabenda V, Mertens R, Fabri V, et al. Adherence to bisphosphonates therapy and hip fracture risk in osteoporotic women. Osteoporos Int 2008;19(6):811-818. 222 Van Wijk BL, Klungel OH, Heerdink ER, et al. Refill persistence with chronic medication assessed from a pharmacy database was influenced by method of calculation. J Clin Epidemiol 2006;59(1):11-17. 223 Nielsen LH, Lokkegaard E, Andreasen AH, et al. Estimating the effect of current, previous and never use of drugs in studies based on prescription registries. Pharmacoepidemiol Drug Saf 2009;18(2):147-153. 224 Suh DC, Thomas SK, Valiyeva E, et al. Drug persistency of two cholinesterase inhibitors: Rivastigmine versus donepezil in elderly patients with Alzheimer's disease. Drugs Aging 2005;22(8):695-707. 225 Halpern R, Agarwal S, Dembek C, et al. Comparison of adherence and persistence among multiple sclerosis patients treated with disease-modifying therapies: A retrospective administrative claims analysis. Patient Prefer Adherence 2011;5:73-84. 226 Mullins CD, Shaya FT, Meng F, et al. Persistence, switching, and discontinuation rates among patients receiving sertraline, paroxetine, and citalopram. Pharmacotherapy 2005;25(5):660-667. 227 Abhishek A, Butt S, Gadsby K, et al. Does temporary discontinuation of anti TNF-alpha agent etanercept affect the therapeutic response at six month [abstract]. Rheumatology (Oxford) 2009;48(Suppl 1):69. 228 Flendrie M. Effectiveness and safety of TNF-alpha blocking  therapy in patients with rheumatoid arthritis [dissertation]. Nijmegen, Netherlands: University of Nijmegen 2009. 229 Harley CR, Frytak JR, Tandon N. Treatment compliance and dosage administration among rheumatoid arthritis patients receiving infliximab, etanercept, or methotrexate. Am J Manag Care 2003;9(6 Suppl):S136-S143. 230 Carter C, Gunnarsson CL, Rizzo J, et al. Economic consequences of under-utilization with tumor necrosis factor inhibitors in rheumatoid arthritis patients [abstract]. Value Health 2011;14(3):A70. 231 Singh JA, Christensen R, Wells GA, et al. Biologics for rheumatoid arthritis: An overview of Cochrane reviews. Cochrane Database Syst Rev 2009;4(4):CD007848. 232 Aletaha D, Stamm T, Kapral T, et al. Survival and effectiveness of leflunomide compared with methotrexate and sulfasalazine in rheumatoid arthritis: A matched observational study. Ann Rheum Dis 2003;62(10):944-951. 233 Ortendahl M, Schettler JD, Fries JF. Factors influencing length of time taking methotrexate in rheumatoid arthritis. J Rheumatol 2000;27(5):1139-1147. 234 Aletaha D, Smolen JS. Effectiveness profiles and dose dependent retention of traditional disease modifying antirheumatic drugs for rheumatoid arthritis. An observational study. J Rheumatol 2002;29(8):1631-1638.    195 235 Luggen ML. Trials are short, disease long: Measuring drug utility beyond clinical trials. J Rheumatol 2004;31(2):205-206. 236 Strangfeld A, Hierse F, Kekow J, et al. Comparative effectiveness of tumour necrosis factor alpha inhibitors in combination with either methotrexate or leflunomide. Ann Rheum Dis 2009;68(12):1856-1862. 237 Kievit W, Adang EM, Fransen J, et al. The effectiveness and medication costs of three anti-tumour necrosis factor alpha agents in the treatment of rheumatoid arthritis from prospective clinical practice data. Ann Rheum Dis 2008;67(9):1229-1234. 238 Fautrel B, Pham T, Mouterde G, et al. Recommendations of the French Society for Rheumatology regarding TNFalpha antagonist therapy in patients with rheumatoid arthritis. Joint Bone Spine 2007;74(6):627-637. 239 Tornero Molina J, Sanmarti Sala R, Rodriguez Valverde V, et al. Update of the consensus statement of the Spanish Society of Rheumatology on the management of biologic therapies in rheumatoid arthritis. Reumatol Clin 2010;6(1):23-36. 240 Brocq O, Millasseau E, Albert C, et al. Effect of discontinuing TNFalpha antagonist therapy in patients with remission of rheumatoid arthritis. Joint Bone Spine 2009;76(4):350-355. 241 Tanaka Y, Takeuchi T, Mimori T, et al. Discontinuation of infliximab after attaining low disease activity in patients with rheumatoid arthritis: RRR (Remission induction by Remicade in RA) study. Ann Rheum Dis 2010;69(7):1286-1291. 242 Nawata M, Saito K, Nakayamada S, et al. Discontinuation of infliximab in rheumatoid arthritis patients in clinical remission. Mod Rheumatol 2008;18(5):460-464. 243 Kievit W, Fransen J, Adang EM, et al. Evaluating guidelines on continuation of anti- tumour necrosis factor treatment after 3 months: Clinical effectiveness and costs of observed care and different alternative strategies. Ann Rheum Dis 2009;68(6):844-849. 244 Furst DE, Keystone EC, Braun J, et al. Updated consensus statement on biological agents for the treatment of rheumatic diseases, 2010. Ann Rheum Dis 2011;70(Suppl 1):i2-i36. 245 Lutt JR, Deodhar A. Rheumatoid arthritis: Strategies in the management of patients showing an inadequate response to TNFalpha antagonists. Drugs 2008;68(5):591-606. 246 Yazici Y, Shi N, John A. Utilization of biologic agents in rheumatoid arthritis in the United States: Analysis of prescribing patterns in 16,752 newly diagnosed patients and patients new to biologic therapy. Bulletin of the NYU Hospital for Joint Diseases 2008;66(2):77-85. 247 Barbaud A, Granel F, Waton J, et al. How to manage hypersensitivity reactions to biological agents?. EJD 2011;21(5):667-674. 248 Combe B. Impact of dosing on treatment with TNF inhibitors: Managing dose adjustment. Clin Exp Rheumatol 2010;28(3 Suppl 59):S13-7.    196 249 The Canadian Agency for Drugs and Technologies in Health (CADTH). Biological response modifier agents for adults  with rheumatoid arthritis (Therapeutic review, clinical and economic review). 2010 Available at: http://www.cadth.ca/media/pdf/TR_RA_Clinical_and_Economic_Overview_e.pdf. Accessed June, 2011. 250 The Canadian Agency for Drugs and Technologies in Health (CADTH) Therapeutic Review Panel. Biological response modifier agents for adults  with rheumatoid arthritis (Final recommendations). 2010 Available at: http://www.cadth.ca/media/pdf/Biologics_for_RA_TRP_Final_Recommendations_e.pdf. Accessed June, 2011. 251 Haraoui B. Canadian Rheumatology Association - Position on the use of biologic agents for the treatment of rheumatoid arthritis (for the subcommittee on biologic agents). 2002 Available at: http://rheum.ca/images/documents/Biologics_in_RA.pdf. Accessed January, 2011. 252 Haraoui B. The CRA position on use of biologic agents for  RA. CRAJ 2002;12(1):5-11. 253 American College of Rheumatology Subcommittee on Rheumatoid Arthritis Guidelines. Guidelines for the management of rheumatoid arthritis: 2002 Update. Arthritis Rheum 2002;46(2):328-346. 254 Curtis JR, Xie F, Chen L, et al. The comparative risk of serious infections among rheumatoid arthritis patients starting or switching biological agents. Ann Rheum Dis 2011;70(8):1401-1406. 255 Thyagarajan V, Norman H, Alexander K, et al. Risk of fatal infection and malignancy related to  use of anti-tumor necrosis factor-alpha biologics  by rheumatoid  arthritis patients [abstract]. Arthritis Rheum 2011;63(Suppl 10):433. 256 Grijalva CG, Kaltenbach L, Arbogast PG, et al. Adherence to disease-modifying antirheumatic drugs and the effects of exposure misclassification on the risk of hospital admission. Arthritis Care Res 2010;62(5):730-734. 257 Setoguchi S, Schneeweiss S, Avorn J, et al. Tumor necrosis factor-alpha antagonist use and heart failure in elderly patients with rheumatoid arthritis. Am Heart J 2008;156(2):336-341. 258 Herrinton LJ, Curtis JR, Chen L, et al. Study design for a comprehensive assessment of biologic safety using multiple healthcare data systems. Pharmacoepidemiol Drug Saf 2011;20(11):1199-1209. 259 Markenson JA, Gibofsky A, Palmer WR, et al. Persistence with anti-tumor necrosis factor therapies in patients with rheumatoid arthritis: Observations from the RADIUS registry. J Rheumatol 2011;38(7):1273-1281. 260 Caspard H, Chan AK, Walker AM. Compliance with a statin treatment in a usual-care setting: Retrospective database analysis over 3 years after treatment initiation in health maintenance organization enrollees with dyslipidemia. Clin Ther 2005;27(10):1639-1646.    197 261 Lee EW, Wei LJ, Amato DA. Cox-type tegression analysis for large numbers of small groups of correlated failure time observations. In: Klein JP, Goel PK, eds. Survival analysis: State of the art. NATO ASI series. Series E, Applied sciences ed. Boston, Massachusetts, U.S.: Kluwer Academic Publishers; 1992:237-247. 262 Gui-shuang Y, Chengcheng L. Statistical analysis of clustered data using SAS® system (NESUG 2006). 2006 Available at: http://www.nesug.org/proceedings/nesug06/an/da01.pdf. Accessed November, 2010. 263 Gharibvand L, Liu L. Analysis of Survival Data with Clustered Events (paper 237-2009). 2009 Available at: http://support.sas.com/resources/papers/proceedings09/237-2009.pdf. Accessed November, 2010. 264 Fraenkel L, Rabidou N, Dhar R. Are rheumatologists' treatment decisions influenced by patients' age? Rheumatology (Oxford) 2006;45(12):1555-1557. 265 Smolen JS, Landewe R, Breedveld FC, et al. EULAR recommendations for the management of rheumatoid arthritis with synthetic and biological disease-modifying antirheumatic drugs. Ann Rheum Dis 2010;69(6):964-975. 266 Maetzel A, Bombardier C, Strand V, et al. How Canadian and US rheumatologists treat moderate or aggressive rheumatoid arthritis: A survey. J Rheumatol 1998;25(12):2331- 2338. 267 Jobanputra P, Wilson J, Douglas K, et al. A survey of British rheumatologists' DMARD preferences for rheumatoid arthritis. Rheumatology (Oxford) 2004;43(2):206-210. 268 Kamal KM, Madhavan SS, Hornsby JA, et al. Use of tumor necrosis factor inhibitors in rheumatoid arthritis: A national survey of practicing United States rheumatologists. Joint Bone Spine 2006;73(6):718-724. 269 Erkan D, Yazici Y, Harrison MJ, et al. Physician treatment preferences in rheumatoid arthritis of differing disease severity and activity: The impact of cost on first-line therapy. Arthritis Care Res 2002;47(3):285-290. 270 Erkan D, Yazici Y, Paget SA. The impact of pharmacoeconomics on the treatment of rheumatoid arthritis: A 3-year longitudinal study. Drug Benefit Trends 2003;15(Suppl C):8-13. 271 Cush JJ. Biological drug use: US perspectives on indications and monitoring. Ann Rheum Dis 2005;64(Suppl 4):iv18-iv23. 272 Curtis JR, Chen L, Harrold LR, et al. Physician preference motivates the use of anti-tumor necrosis factor therapy independent of clinical disease activity. Arthritis Care Res 2010;62(1):101-107. 273 Curtis JR, Chen L, Harrold LR, et al. Physician preference motivates use of anti-TNF therapy independent of clinical disease activity [abstract]. Arthritis Rheum 2009;60(Suppl 10):1597. 274 Schneeweiss S, Glynn RJ, Avorn J, et al. A Medicare database review found that physician preferences increasingly outweighed patient characteristics as determinants of first-time prescriptions for COX-2 inhibitors. J Clin Epidemiol 2005;58(1):98-102.    198 275 Solomon DH, Schneeweiss S, Glynn RJ, et al. Determinants of selective cyclooxygenase-2 inhibitor prescribing: Are patient or physician characteristics more important? Am J Med 2003;115(9):715-720. 276 Stedman MR, Brookhart MA. Bayesian prescribing model for COX-2 selective NSAIDS [abstract]. Pharmacoepidemiol Drug Saf 2009;18(Suppl S1):S31. 277 Baser O, Wang L, Xie L, et al. Deriving doctors' prescribing patterns from health care claims: An instrumental variable analysis [abstract]. Value Health 2010;13(7):A425. 278 Yazici Y, Simsek I. Treatment options for rheumatoid arthritis beyond TNF-alpha inhibitors. Expert Rev Clin Pharmacol 2010;3(5):663-666. 279 Dudler J, Moller B, Michel BA, et al. Biologics in rheumatoid arthritis (RA) – Recommendations for Swiss practice. Swiss Med Wkly 2011;(w13189). 280 Lu TY, Hill C. Managing patients taking tumour necrosis factor inhibitors. Australian Prescriber 2006;29(3):67-70. 281 Gall EP, Caperton EM, McComb JE, et al. Clinical comparison of ibuprofen, fenoprofen calcium, naproxen and tolmetin sodium in rheumatoid arthritis. J Rheumatol 1982;9(3):402-407. 282 Wasner C, Britton MC, Kraines RG, et al. Nonsteroidal anti-inflammatory agents in rheumatoid arthritis and ankylosing spondylitis. JAMA 1981;246(19):2168-2172. 283 Choquette D, Shaikh SA, Abu-Hakima M. Rheumatoid Arthritis Patient Preference Survey (RAPPS) (Presentation number: 965). 2006 Available at: http://acr.confex.com/acr/2006/webprogram/Paper5515.html. Accessed December, 2011. 284 Tandon N, Bolge S, Annunziata K. Satisfaction, preference and reasons for biologic treatment discontinuation in patients with rheumatoid arthritis [abstract]. Value Health 2011;14(3):A134. 285 Williams EL, Edwards CJ. Patient preferences in choosing anti-TNF therapies-R1 [letter to the editor]. Rheumatology (Oxford) 2006;45(12):1575-1576. 286 Fraenkel L, Bogardus ST, Concato J, et al. Patient preferences for treatment of rheumatoid arthritis. Ann Rheum Dis 2004;63(11):1372-1378. 287 van Hulst LT, Kievit W, van Bommel R, et al. Rheumatoid arthritis patients and rheumatologists approach the decision to escalate care differently: Results of a maximum difference scaling experiment. Arthritis Care Res 2011;63(10):1407-1414. 288 Kievit W, van Hulst LT, van Riel P, et al. Factors that influence rheumatologists' decisions to escalate care in rheumatoid arthritis: Results from a choice-based conjoint analysis. Arthritis Care Res 2010;62(6):842-847. 289 Curtis JR, Shan Y, Kremer JM, et al. Disease activity, calendar year, and physician preference are associated with switching of biologic agents for reasons of inadequate efficacy [abstract]. Arthritis Rheum 2010;62(Suppl 10):379. 290 Mark TL, Swait J. Using stated preference and revealed preference modeling to evaluate prescribing decisions. Health Econ 2004;13(6):563-573.    199 291 Festinger L. A theory of cognitive dissonance. Evanston, Ill.: Row, Peterson 1957:291. 292 Visser PS, Cooper J. Dissonance and attitude change: A matter of measurement, in Chapter 10: Attitude change. In: Hogg MA, Cooper J, eds. The Sage handbook of social psychology Thousand Oaks, California, U.S.: Sage; 2003:220. 293 Brookhart MA, Avorn J, Katz JN, et al. Gaps in treatment among users of osteoporosis medications: The dynamics of noncompliance. Am J Med 2007;120(3):251-256. 294 Arnett FC, Edworthy SM, Bloch DA, et al. The American Rheumatism Association 1987 revised criteria for the classification of rheumatoid arthritis. Arthritis Rheum 1988;31(3):315-324. 295 Pincus T, Callahan LF. The 'side effects' of rheumatoid arthritis: Joint destruction, disability and early mortality [abstract]. Br J Rheumatol 1993;32(Suppl 1):28-37. 296 Alamanos Y, Voulgari PV, Drosos AA. Incidence and prevalence of rheumatoid arthritis, based on the 1987 American College of Rheumatology criteria: A systematic review. Semin Arthritis Rheum 2006;36(3):182-188. 297 Shapira Y, Agmon-Levin N, Shoenfeld Y. Geoepidemiology of autoimmune rheumatic diseases. Nat Rev Rheumatol 2010;6(8):468-476. 298 Morse A, Groom C, Dixon P, et al. Services for people with rheumatoid arthritis – International comparisons [Report No HC 823 Session 2008-2009]. 2009 Available at: http://www.nao.org.uk/publications/0809/services_for_people_with_rheum.aspx. Accessed May, 2011. 299 Lacaille D, Anis A, Guh D, et al. Assessing the quality of care for RA at a population level (abstract 1687) [abstract]. J Rheumatol 2002;30(8):S1883. 300 Lacaille D. Gaps in care for rheumatoid arthritis: The BC story. 2003 Available at: http://www.arthritisresearch.ca/top-menu-research-current/78- gapsincare.html?catid=63%3Ara-current-research. Accessed September, 2011. 301 BC Stats. Census population of B.C. & Canada, 1871 to 2006. 2006 Available at: http://www.bcstats.gov.bc.ca/data/cen06/c2006bc.asp. Accessed September, 2011. 302 Lacaille D. Databases, registries and you: The Canadian experience. What we learned from administrative databases [CRA conference slides]. 2010 Available at: http://rheum.ca/Resources/CRA-Mtgs/2010/Slides/Lacaille.pdf. Accessed August, 2011. 303 Gabriel SE, Crowson CS, Maradit Kremers H, et al. The rising incidence of rheumatoid arthritis [abstract]. Arthritis Rheum 2008;58:S453. 304 Myasoedova E, Crowson CS, Kremers HM, et al. Is the incidence of rheumatoid arthritis rising?: Results from Olmsted County, Minnesota, 1955-2007. Arthritis Rheum 2010;62(6):1576-1582. 305 Lawrence RC, Helmick CG, Arnett FC, et al. Estimates of the prevalence of arthritis and selected musculoskeletal disorders in the United States. Arthritis Rheum 1998;41(5):778- 799.    200 306 Imanaka T, Shichikawa K, Inoue K, et al. Increase in age at onset of rheumatoid arthritis in Japan over a 30 year period. Ann Rheum Dis 1997;56(5):313-316. 307 Gabriel SE, Crowson CS, O'Fallon WM. The epidemiology of rheumatoid arthritis in Rochester, Minnesota, 1955-1985. Arthritis Rheum 1999;42(3):415-420. 308 Marra CA. Outcome measures in economic evaluations of rheumatoid arthritis [Dissertation]. Vancouver, British Columbia, Canada: The University of British Columbia 2004. 309 Maetzel A, Li LC, Pencharz J, et al. The economic burden associated with osteoarthritis, rheumatoid arthritis, and hypertension: A comparative study. Ann Rheum Dis 2004;63(4):395-401. 310 Fautrel B, Clarke AE, Guillemin F, et al. Costs of rheumatoid arthritis: New estimates from the human capital method and comparison to the willingness-to-pay method. Med Decis Making 2007;27(2):138-150. 311 Grosser T, Smyth E, FitzGerald GA. Disease-modifying anti-rheuamtic drugs, in Chapter 34. Anti-inflammatory, antipyretic, and analgesic agents; Pharmacotherapy of gout. In: Goodman LS, Brunton LL, Chabner B, et al., eds. Goodman & Gilman's pharmacological basis of therapeutics. 12th ed. New York, New York, U.S.: McGraw-Hill; 2011. 312 Plosker GL, Croom KF. Sulfasalazine: A review of its use in the management of rheumatoid arthritis. Drugs 2005;65(13):1825-1849. 313 Kitahara K, Kawai S. Cyclosporine and tacrolimus for the treatment of rheumatoid arthritis. Curr Opin Rheumatol 2007;19(3):238-245. 314 Maltzman JS, Koretzky GA. Azathioprine: Old drug, new actions. J Clin Invest 2003;111(8):1122-1124. 315 Segal R, Yaron M, Tartakovsky B. Methotrexate: Mechanism of action in rheumatoid arthritis. Semin Arthritis Rheum 1990;20(3):190-200. 316 Breedveld FC, Dayer JM. Leflunomide: Mode of action in the treatment of rheumatoid arthritis. Ann Rheum Dis 2000;59(11):841-849. 317 Fox RI. Mechanism of action of leflunomide in rheumatoid arthritis. J Rheumatol 1998;53(Suppl):20-26. 318 Fox RI. Mechanism of action of hydroxychloroquine as an antirheumatic drug. Semin Arthritis Rheum 1993;23(2 Suppl 1):82-91. 319 Burger D, Dayer JM, Palmer G, et al. Is IL-1 a good therapeutic target in the treatment of arthritis? Best Pract Res Clin Rheumatol 2006;20(5):879-896. 320 Gibiansky L, Frey N. Linking interleukin-6 receptor blockade with tocilizumab and its hematological effects using a modeling approach. J Pharmacokinet Pharmacodyn 2012;39(1):5-16. 321 Taylor RP, Lindorfer MA. Drug insight: The mechanism of action of rituximab in autoimmune disease – The immune complex decoy hypothesis. Nat Clin Pract Rheumatol 2007;3(2):86-95.    201 322 Goeb V, Buch MH, Vital EM, et al. Costimulation blockade in rheumatic diseases: Where we are? Curr Opin Rheumatol 2009;21(3):244-250. 323 O'Dell JR. Therapeutic strategies for rheumatoid arthritis. N Engl J Med 2004;350(25):2591-2602. 324 Lard LR, Visser H, Speyer I, et al. Early versus delayed treatment in patients with recent- onset rheumatoid arthritis: Comparison of two cohorts who received different treatment strategies. Am J Med 2001;111(6):446-451. 325 Egsmose C, Lund B, Borg G, et al. Patients with rheumatoid arthritis benefit from early 2nd line therapy: 5 year followup of a prospective double blind placebo controlled study. J Rheumatol 1995;22(12):2208-2213. 326 Tsakonas E, Fitzgerald AA, Fitzcharles MA, et al. Consequences of delayed therapy with second-line agents in rheumatoid arthritis: A 3 year followup on the hydroxychloroquine in early rheumatoid arthritis (HERA) study. J Rheumatol 2000;27(3):623-629. 327 Finckh A, Liang MH, van Herckenrode CM, et al. Long-term impact of early treatment on radiographic progression in rheumatoid arthritis: A meta-analysis. Arthritis Rheum 2006;55(6):864-872. 328 Katchamart W, Trudeau J, Phumethum V, et al. Methotrexate monotherapy versus methotrexate combination therapy with non-biologic disease modifying anti-rheumatic drugs for rheumatoid arthritis. Cochrane Database Syst Rev 2010;4(4):CD008495. 329 Donahue KE, Gartlehner G, Jonas DE, et al. Systematic review: Comparative effectiveness and harms of disease-modifying medications for rheumatoid arthritis. Ann Intern Med 2008;148(2):124-134. 330 Saag KG, Gim GT, Patkar NM, et al. American College of Rheumatology 2008 recommendations for the use of nonbiologic and biologic disease-modifying antirheumatic drugs in rheumatoid arthritis. Arthritis Care Res 2008;59(6):762-784. 331 Lipsky PE, van der Heijde DM, St Clair EW, et al. Infliximab and methotrexate in the treatment of rheumatoid arthritis. Anti-Tumor Necrosis Factor Trial in Rheumatoid Arthritis with Concomitant Therapy Study Group. N Engl J Med 2000;343(22):1594-1602. 332 Sokka T. Radiographic scoring in rheumatoid arthritis: A short introduction to the methods. Bull NYU Hosp Jt Dis 2008;66(2):166-168. 333 Jacobsson LT, Turesson C, Nilsson JA, et al. Treatment with TNF blockers and mortality risk in patients with rheumatoid arthritis. Ann Rheum Dis 2007;66(5):670-675. 334 Carmona L, Descalzo MA, Perez-Pampin E, et al. All-cause and cause-specific mortality in rheumatoid arthritis are not greater than expected when treated with tumour necrosis factor antagonists. Ann Rheum Dis 2007;66(7):880-885. 335 Olofsson T, Englund M, Saxne T, et al. Decrease in sick leave among patients with rheumatoid arthritis in the first 12 months after start of treatment with tumour necrosis factor antagonists: A population-based controlled cohort study. Ann Rheum Dis 2010;69(12):2131-2136.    202 336 Augustsson J, Neovius M, Cullinane-Carli C, et al. Patients with rheumatoid arthritis treated with tumour necrosis factor antagonists increase their participation in the workforce: Potential for significant long-term indirect cost gains (data from a population- based registry). Ann Rheum Dis 2010;69(1):126-131. 337 Sandhu RS, Treharne GJ, Douglas KM, et al. The impact of anti-tumour necrosis factor therapy for rheumatoid arthritis on the use of other drugs and hospital resources in a pragmatic setting. Muscoskel Care 2006;4(4):204-222. 338 Bongartz T, Sutton AJ, Sweeting MJ, et al. Anti-TNF antibody therapy in rheumatoid arthritis and the risk of serious infections and malignancies: Systematic review and meta- analysis of rare harmful effects in randomized controlled trials. JAMA 2006;295(19):2275- 2285. 339 Smitten AL, Simon TA, Hochberg MC, et al. A meta-analysis of the incidence of malignancy in adult patients with rheumatoid arthritis. Arthritis Res Ther 2008;10(2):R45. 340 Robinson WH, Genovese MC, Moreland LW. Demyelinating and neurologic events reported in association with tumor necrosis factor alpha antagonism: By what mechanisms could tumor necrosis factor alpha antagonists improve rheumatoid arthritis but exacerbate multiple sclerosis? Arthritis Rheum 2001;44(9):1977-1983. 341 Wajant H, Pfizenmaier K, Scheurich P. Tumor necrosis factor signaling. Cell Death Differ 2003;10(1):45-65. 342 Abbas AK, Lichtman AH. Effector mechanisms of cell-mediated immunity. In: Abbas AK, Lichtman AH, eds. Basic immunology: Functions and disorders of the immune system. 3rd ed. Philadelphia, Pennsylvania, U.S.: Saunders/Elsevier; 2009:113-129. 343 Janeway C, Murphy KP, Travers P, Walport M. Chapter 8: T cell-mediated immunity. In: Janeway C, Murphy KP, Travers P, et al., eds. Janeway's immunobiology. 7th ed. New York, New York, U.S.: Garland Science; 2008:887-378. 344 Arnett HA, Mason J, Marino M, et al. TNF alpha promotes proliferation of oligodendrocyte progenitors and remyelination. Nat Neurosci 2001;4(11):1116-1122. 345 Kury P, Stoll G, Muller HW. Molecular mechanisms of cellular interactions in peripheral nerve regeneration. Curr Opin Neurol 2001;14(5):635-639. 346 Collins RA, Grounds MD. The role of tumor necrosis factor-alpha (TNF-alpha) in skeletal muscle regeneration. Studies in TNF-alpha(-/-) and TNF-alpha(-/-)/LT-alpha(-/-) mice. J Histochem Cytochem 2001;49(8):989-1001. 347 Tracey KJ, Cerami A. Tumor necrosis factor: An updated review of its biology. Crit Care Med 1993;21(10 Suppl):S415-S422. 348 Perez C, Albert I, DeFay K, et al. A nonsecretable cell surface mutant of tumor necrosis factor (TNF) kills by cell-to-cell contact. Cell 1990;63(2):251-258. 349 Horiuchi T, Mitoma H, Harashima S, et al. Transmembrane TNF-alpha: Structure, function and interaction with anti-TNF agents. Rheumatology (Oxford) 2010;49(7):1215- 1228. 350 Barnes PJ. Nuclear factor-kappa B. Int J Biochem Cell Biol 1997;29(6):867-870.    203 351 Dayer JM, Beutler B, Cerami A. Cachectin/tumor necrosis factor stimulates collagenase and prostaglandin E2 production by human synovial cells and dermal fibroblasts. J Exp Med 1985;162(6):2163-2168. 352 Bertolini DR, Nedwin GE, Bringman TS, et al. Stimulation of bone resorption and inhibition of bone formation in vitro by human tumour necrosis factors. Nature 1986;319(6053):516-518. 353 Brennan FM, Chantry D, Jackson AM, et al. Cytokine production in culture by cells isolated from the synovial membrane. J Autoimmun 1989;2(Suppl 1):177-186. 354 Kavanaugh AF, Schulze-Koops H, Davis LS, et al. Repeat treatment of rheumatoid arthritis patients with a murine anti- intercellular adhesion molecule 1 monoclonal antibody. Arthritis Rheum 1997;40(5):849-853. 355 Borzi RM, Arfilli L, Focherini MC, et al. Circulating tumor necrosis factor alpha in rheumatoid arthritis. Boll Soc Ital Biol Sper 1993;69(1):39-43. 356 Saxne T, Palladino J, M.a, et al. Detection of tumor necrosis factor alpha but not tumor necrosis factor beta in rheumatoid arthritis synovial fluid and serum. Arthritis Rheum 1988;31(8):1041-1045. 357 Neidel J, Schulze M, Lindschau J. Association between degree of bone-erosion and synovial fluid-levels of tumor necrosis factor alpha in the knee-joints of patients with rheumatoid arthritis. Inflamm Res 1995;44(5):217-221. 358 Martinez A, Fernandez-Arquero M, Pascual-Salcedo D, et al. Primary association of tumor necrosis factor-region genetic markers with susceptibility to rheumatoid arthritis. Arthritis Rheum 2000;43(6):1366-1370. 359 Lee YH, Ji JD, Song GG. Tumor necrosis factor-alpha promoter -308 A/G polymorphism and rheumatoid arthritis susceptibility: A metaanalysis. J Rheumatol 2007;34(1):43-49. 360 Rodriguez-Carreon AA, Zuniga J, Hernandez-Pacheco G, et al. Tumor necrosis factor- alpha -308 promoter polymorphism contributes independently to HLA alleles in the severity of rheumatoid arthritis in Mexicans. J Autoimmun 2005;24(1):63-68. 361 Brennan FM, McInnes IB. Evidence that cytokines play a role in rheumatoid arthritis. J Clin Invest 2008;118(11):3537-3545. 362 Nestorov I. Clinical pharmacokinetics of tumor necrosis factor antagonists. J Rheumatol 2005;32(74):13-18. 363 Dirks NL, Meibohm B. Population pharmacokinetics of therapeutic monoclonal antibodies. Clin Pharmacokinet 2010;49(10):633-659. 364 Jit M, Henderson B, Stevens M, et al. TNF-alpha neutralization in cytokine-driven diseases: A mathematical model to account for therapeutic success in rheumatoid arthritis but therapeutic failure in systemic inflammatory response syndrome. Rheumatology (Oxford) 2005;44(3):323-331. 365 Wagner C, Mace K, DeWoody K, et al. Infliximab treatment benefits correlate with pharmacodynamic parameters in Crohn’s disease patients [abstract]. Digestion 1998;59(Suppl 3):124-125.    204 366 Mohler KM, Torrance DS, Smith CA, et al. Soluble tumor necrosis factor (TNF) receptors are effective therapeutic agents in lethal endotoxemia and function simultaneously as both TNF carriers and TNF antagonists. J Immunol 1993;151(3):1548-1561. 367 Suffredini AF, Reda D, Banks SM, et al. Effects of recombinant dimeric TNF receptor on human inflammatory responses following intravenous endotoxin administration. J Immunol 1995;155(10):5038-5045. 368 Barrera P, Joosten LA, den Broeder AA, et al. Effects of treatment with a fully human anti-tumour necrosis factor alpha monoclonal antibody on the local and systemic homeostasis of interleukin 1 and TNFalpha in patients with rheumatoid arthritis. Ann Rheum Dis 2001;60(7):660-669. 369 Charles P, Elliott MJ, Davis D, et al. Regulation of cytokines, cytokine inhibitors, and acute-phase proteins following anti-TNF-alpha therapy in rheumatoid arthritis. J Immunol 1999;163(3):1521-1528. 370 Scallon B, Cai A, Solowski N, et al. Binding and functional comparisons of two types of tumor necrosis factor antagonists. J Pharmacol Exp Ther 2002;301(2):418-426. 371 Santora LC, Kaymakcalan Z, Sakorafas P, et al. Characterization of noncovalent complexes of recombinant human monoclonal antibody and antigen using cation exchange, size exclusion chromatography, and BIAcore. Anal Biochem 2001;299(2):119- 129. 372 Kaymakcalan Z, Kalghatgi L, Xiong L. Differential TNF-neutralizing potencies of adalimumab, etanercept and infliximab [abstract]. Clin Immunol 2006;119(Suppl):S77- S79. 373 Tumanov AV, Christiansen PA, Fu YX. The role of lymphotoxin receptor signaling in diseases. Curr Mol Med 2007;7(6):567-578. 374 Spencer-Green G. Etanercept (Enbrel): Update on therapeutic use. Ann Rheum Dis 2000;59(Suppl 1):i46-i49. 375 Shen C, Assche GV, Colpaert S, et al. Adalimumab induces apoptosis of human monocytes: A comparative study with infliximab and etanercept. Aliment Pharmacol Ther 2005;21(3):251-258. 376 Catrina AI, Trollmo C, af Klint E, et al. Evidence that anti-tumor necrosis factor therapy with both etanercept and infliximab induces apoptosis in macrophages, but not lymphocytes, in rheumatoid arthritis joints: Extended report. Arthritis Rheum 2005;52(1):61-72. 377 Hwang WY, Foote J. Immunogenicity of engineered antibodies. Methods 2005;36(1):3-10. 378 eCPS (online Compendium of Pharmaceuticals and Specialties). Humira® (adalimumab). 2009. Accessed June, 2009. 379 eCPS (online Compendium of Pharmaceuticals and Specialties). Enbrel® (Etanercept). 2009. Accessed June, 2009.    205 380 Svenson M, Geborek P, Saxne T, et al. Monitoring patients treated with anti-TNF-alpha biopharmaceuticals: Assessing serum infliximab and anti-infliximab antibodies. Rheumatology (Oxford) 2007;46(12):1828-1834. 381 Wolbink GJ, Vis M, Lems W, et al. Development of antiinfliximab antibodies and relationship to clinical response in patients with rheumatoid arthritis. Arthritis Rheum 2006;54(3):711-715. 382 van den Bemt BJ, den Broeder AA, Snijders GF, et al. Sustained effect after lowering high-dose infliximab in patients with rheumatoid arthritis: A prospective dose titration study. Ann Rheum Dis 2008;67(12):1697-1701. 383 Radstake TR, Svenson M, Eijsbouts AM, et al. Formation of antibodies against infliximab and adalimumab strongly correlates with functional drug levels and clinical responses in rheumatoid arthritis. Ann Rheum Dis 2009;68(11):1739-1745. 384 van der Laken CJ, Voskuyl AE, Roos JC, et al. Imaging and serum analysis of immune complex formation of radiolabelled infliximab and anti-infliximab in responders and non- responders to therapy for rheumatoid arthritis. Ann Rheum Dis 2007;66(2):253-256. 385 van der Bijl AE, Breedveld FC, Antoni CE, et al. An open-label pilot study of the effectiveness of adalimumab in patients with rheumatoid arthritis and previous infliximab treatment: Relationship to reasons for failure and anti-infliximab antibody status. Clin Rheumatol 2008;27(8):1021-1028. 386 Vultaggio A, Matucci A, Nencini F, et al. Anti-infliximab IgE and non-IgE antibodies and induction of infusion-related severe anaphylactic reactions. Allergy 2010;65(5):657-661. 387 Bartelds GM, Krieckaert CL, Nurmohamed MT, et al. Immunogenicity in a 3-year follow- up cohort of adalimumab treated rheumatoid arthritis patients [abstract]. Arthritis Rheum 2010;62(Suppl 10):1445. 388 Weinblatt ME, Keystone EC, Furst DE, et al. Adalimumab, a fully human anti-tumor necrosis factor alpha monoclonal antibody, for the treatment of rheumatoid arthritis in patients taking concomitant methotrexate: The ARMADA trial. Arthritis Rheum 2003;48(1):35-45. 389 Bartelds GM, Wijbrandts CA, Nurmohamed MT, et al. Clinical response to adalimumab: Relationship to anti-adalimumab antibodies and serum adalimumab concentrations in rheumatoid arthritis. Ann Rheum Dis 2007;66(7):921-926. 390 van de Putte LB, Atkins C, Malaise M, et al. Efficacy and safety of adalimumab as monotherapy in patients with rheumatoid arthritis for whom previous disease modifying antirheumatic drug treatment has failed. Ann Rheum Dis 2004;63(5):508-516. 391 Bender NK, Heilig CE, Droll B, et al. Immunogenicity, efficacy and adverse events of adalimumab in RA patients. Rheumatol Int 2007;27(3):269-274. 392 Dore RK, Mathews S, Schechtman J, et al. The immunogenicity, safety, and efficacy of etanercept liquid administered once weekly in patients with rheumatoid arthritis. Clin Exp Rheumatol 2007;25(1):40-46.    206 393 The electronic Medicines Compendium (eMC) UK. Remicade 100 mg powder for concentrate for solution for infusion. 2011 Available at: http://www.medicines.org.uk/emc/medicine/3236. Accessed July, 2011. 394 Zhou H. Clinical pharmacokinetics of etanercept: A fully humanized soluble recombinant tumor necrosis factor receptor fusion protein. J Clin Pharmacol 2005;45(5):490-497. 395 Kavanaugh A, St Clair EW, McCune WJ, et al. Chimeric anti-tumor necrosis factor-alpha monoclonal antibody treatment of patients with rheumatoid arthritis receiving methotrexate therapy. J Rheumatol 2000;27(4):841-850. 396 Paserchia LA. Clinical pharmacology review of BLA 99-0128,  Remicade (supplement) [report no BLA99-0128]. 1999 Available at: http://www.fda.gov/Drugs/DevelopmentApprovalProcess/HowDrugsareDevelopedandApp roved/ApprovalApplications/TherapeuticBiologicApplications/ucm093335.htm. Accessed May, 2011. 397 Mahmood I. Clinical pharmacology and biopharmaceutics review – Adalimumab (40 mg subcutaneous administration) [report No 125057/0]. 2002 Available at: http://www.fda.gov/downloads/Drugs/DevelopmentApprovalProcess/HowDrugsareDevelo pedandApproved/ApprovalApplications/TherapeuticBiologicApplications/ucm092770.pdf . Accessed May, 2011. 398 Weisman MH, Moreland LW, Furst DE, et al. Efficacy, pharmacokinetic, and safety assessment of adalimumab, a fully human anti-tumor necrosis factor-alpha monoclonal antibody, in adults with rheumatoid arthritis receiving concomitant methotrexate: A pilot study. Clin Ther 2003;25(6):1700-1721. 399 Green MD. Nonclinical pharmacology and toxicology review  of BLA 98-0286 – Etanercept [report]. 1999 Available at: http://www.fda.gov/downloads/Drugs/DevelopmentApprovalProcess/HowDrugsareDevelo pedandApproved/ApprovalApplications/TherapeuticBiologicApplications/ucm088697.pdf . Accessed June, 2011. 400 Kremer J, Spencer-Green G, Hanna R, et al. Enbrel® (etanrecept) pharmacokinetics in patients with  rheumatoid arthritis [abstract]. Arthritis Rheum 2000;43:S976. 401 Keystone EC, Schiff MH, Kremer JM, et al. Once-weekly administration of 50 mg etanercept in patients with active rheumatoid arthritis: Results of a multicenter, randomized, double-blind, placebo-controlled trial. Arthritis Rheum 2004;50(2):353-363. 402 St.Clair EW, Wagner CL, Fasanmade AA, et al. The relationship of serum infliximab concentrations to clinical improvement in rheumatoid arthritis: Results from ATTRACT, a multicenter, randomized, double-blind, placebo-controlled trial. Arthritis Rheum 2002;46(6):1451-1459. 403 Granneman RG, Zhang Y, Noertersheuser PA, et al. Pharmacokinetic/ pharmacodynamic (PK/PD) relationships of adalimumab (Humira, Abbott) in rheumatoid arthritis (RA) patients during phase II/III clinical trials [abstract]. Arthritis Rheum 2003;48(Suppl):S140- S141. 404 Stanley K. Design of randomized controlled trials. Circulation 2007;115(9):1164-1169.    207 405 Felson DT, Anderson JJ, Boers M, et al. The American College of Rheumatology preliminary core set of disease activity measures for rheumatoid arthritis clinical trials. The committee on outcome measures in rheumatoid arthritis clinical trials. Arthritis Rheum 1993;36(6):729-740. 406 Wernick R, Campbell SM. Update in rheumatology. Ann Intern Med 2000;132(2):125- 133. 407 Gartlehner G, Hansen RA, Jonas BL, et al. The comparative efficacy and safety of biologics for the treatment of rheumatoid arthritis: A systematic review and metaanalysis. J Rheumatol 2006;33(12):2398-2408. 408 Nixon RM, Bansback NJ, Brennan A. Using mixed treatment comparisons and meta- regression to perform indirect comparison to estimate the efficacy of biologic treatments in rheumatoid arthritis. Stat Med 2007;26(6):1237-1254. 409 Alonso-Ruiz A, Pijoan JI, Ansuategui E, et al. Tumor necrosis factor alpha drugs in rheumatoid arthritis: Systematic review and metaanalysis of efficacy and safety. BMC Musculoskelet Disord 2008;9:52. 410 Hochberg MC, Tracy JK, Hawkins-Holt M, et al. Comparison of the efficacy of the tumour necrosis factor alpha blocking agents adalimumab, etanercept, and infliximab when added to methotrexate in patients with active rheumatoid arthritis. Ann Rheum Dis 2003;62(Suppl 2):ii13-ii16. 411 Wiens A, Venson R, Correr CJ, et al. Meta-analysis of the efficacy and safety of adalimumab, etanercept, and infliximab for the treatment of rheumatoid arthritis. Pharmacotherapy 2010;30(4):339-353. 412 Ingham M, Carter C. Meta-analysis and biologics comparing apples to apples: An example using radiographic outcomes in rheumatoid arthritis [abstract]. Arthritis Rheum 2010;62(Suppl 10):76. 413 Singh JA, Wells GA, Christensen R, et al. Adverse effects of biologics: A network meta- analysis and Cochrane overview. Cochrane Database Syst Rev 2011;(2):CD008794. 414 Devine EB, Alfonso-Cristancho R, Sullivan SD. Effectiveness of biologic therapies for rheumatoid arthritis: An indirect comparisons approach. Pharmacotherapy 2011;31(1):39- 51. 415 Herland K, Akselsen JP, Skjønsberg OH, et al. How representative are clinical study patients with asthma or COPD for a larger “real life” population of patients with obstructive lung disease? Respir Med 2005;99(1):11-19. 416 Hordijk-Trion M, Lenzen M, Wijns W, et al. Patients enrolled in coronary intervention trials are not representative of patients in clinical practice: Results from the Euro Heart Survey on Coronary Revascularization. Eur Heart J 2006;27(6):671-678. 417 Engen CN, Engebretsen L, Årøen A. Knee cartilage defect patients enrolled in randomized  controlled trials are not representative of patients in orthopedic practice. Cartilage 2010;1(4):312-319.    208 418 Maasland L, van Oostenbrugge RJ, Franke CF, et al. Patients enrolled in large randomized clinical trials of antiplatelet treatment for prevention after transient ischemic attack or ischemic stroke are not representative of patients in clinical practice: The Netherlands Stroke Survey. Stroke 2009;40(8):2662-2668. 419 Fortin M, Dionne J, Pinho G, et al. Randomized controlled trials: Do they have external validity for patients with multiple comorbidities? Ann Fam Med 2006;4(2):104-108. 420 Dans AL, Dans LF, Guyatt GH, et al. Users' guides to the medical literature: XIV. How to decide on the applicability of clinical trial results to your patient. Evidence-Based Medicine Working Group. JAMA 1998;279(7):545-549. 421 Nadel MV. Problems in implementing policy on women in study populations [report no T- HRD-90-38]. 1990 Available at: http://161.203.16.4/d48t13/141601.pdf. Accessed October, 2011. 422 Lee PY, Alexander KP, Hammill BG, et al. Representation of elderly persons and women in published randomized trials of acute coronary syndromes. JAMA 2001;286(6):708-713. 423 Healy B. The Yentl syndrome. N Engl J Med 1991;325(4):274-276. 424 Heiat A, Gross CP, Krumholz HM. Representation of the elderly, women, and minorities in heart failure clinical trials. Arch Intern Med 2002;162(15):1682-1688. 425 Hoel AW, Kayssi A, Brahmanandam S, et al. Under-representation of women and ethnic minorities in vascular surgery randomized controlled trials. J Vasc Surg 2009;50(2):349- 354. 426 Robinson D, Woerner MG, Pollack S, et al. Subject selection biases in clinical trials: Data from a multicenter schizophrenia treatment study. J Clin Psychopharmacol 1996;16(2):170-176. 427 Chaves AC, Seeman MV. Sex selection bias in schizophrenia antipsychotic trials. J Clin Psychopharmacol 2006;26(5):489-494. 428 Pincus T, Sokka T. Clinical trials in rheumatic diseases: Designs and limitations. Rheum Dis Clin North Am 2004;30(4):701-724. 429 Berthelot JM, Benoist-Gerard S, le Goff B, et al. Outcome and safety of TNFalpha antagonist therapy in 475 consecutive outpatients (with rheumatoid arthritis or spondyloarthropathies) treated by a single physician according to their eligibility for clinical trials. Joint Bone Spine 2010;77(6):564-569. 430 van Riel PL, van Gestel AM, van de Putte LB. Development and validation of response criteria in rheumatoid arthritis: Steps towards an international consensus on prognostic markers. Br J Rheumatol 1996;35(Suppl 2):4-7. 431 Book C, Saxne T, Jacobsson LT. Prediction of mortality in rheumatoid arthritis based on disease activity markers. J Rheumatol 2005;32(3):430-434. 432 Chehata JC, Hassell AB, Clarke SA, et al. Mortality in rheumatoid arthritis: Relationship to single and composite measures of disease activity. Rheumatology (Oxford) 2001;40(4):447-452.    209 433 Callahan LF, Pincus T, Huston JW,3rd, et al. Measures of activity and damage in rheumatoid arthritis: Depiction of changes and prediction of mortality over five years. Arthritis Care Res 1997;10(6):381-394. 434 Pincus T, Brooks RH, Callahan LF. Prediction of long-term mortality in patients with rheumatoid arthritis according to simple questionnaire and joint count measures. Ann Intern Med 1994;120(1):26-34. 435 Wolfe F, Michaud K, Dewitt EM. Why results of clinical trials and observational studies of antitumour necrosis factor (anti-TNF) therapy differ: Methodological and interpretive issues. Ann Rheum Dis 2004;63(Suppl 2):ii13-ii17. 436 Wolfe F, Michaud K, Gefeller O, et al. Predicting mortality in patients with rheumatoid arthritis. Arthritis Rheum 2003;48(6):1530-1542. 437 Andrade S. Compliance in the real world. Value Health 1998;1(3):171-173. 438 Lunt M, Watson KD, Dixon WG, et al. No evidence of association between anti-tumor necrosis factor treatment and mortality in patients with rheumatoid arthritis: Results from the British Society for Rheumatology Biologics Register. Arthritis Rheum 2010;62(11):3145-3153. 439 Fransen J, van Riel PL. The Disease Activity Score and the EULAR response criteria. Rheum Dis Clin North Am 2009;35(4):745-57,-vii-viii. 440 van der Heijde DM, van 't Hof MA, van Riel PL, et al. Judging disease activity in clinical practice in rheumatoid arthritis: First step in the development of a disease activity score. Ann Rheum Dis 1990;49(11):916-920. 441 van Gestel AM, Prevoo ML, van 't Hof MA, et al. Development and validation of the European League Against Rheumatism response criteria for rheumatoid arthritis. Comparison with the preliminary American College of Rheumatology and the World Health Organization/International League Against Rheumatism Criteria. Arthritis Rheum 1996;39(1):34-40. 442 National Institute for Health and Clinical Excellence (NICE). Rheumatoid arthritis - The management of rheumatoid arthritis in adults (NICE clinical guideline 79). 2009 Available at: http://www.nice.org.uk/nicemedia/pdf/CG79NICEGuideline.pdf. Accessed November, 2010. 443 Valesini G, Montecucco C, Cutolo M. Recommendations for the use of biologic (TNF- alpha blocking) agents in the treatment of rheumatoid arthritis in Italy. Clin Exp Rheumatol 2006;24(4):413-423. 444 Furst DE, Breedveld FC, Kalden JR, et al. Updated consensus statement on biological agents for the treatment of rheumatoid arthritis and other immune mediated inflammatory diseases (May 2003). Ann Rheum Dis 2003;62(Suppl 2):ii2-ii9. 445 Haraoui B. Canadian consensus on the use of TNF antagonists in the treatment of RA. CRAJ 2000;10(2):10-12-12.    210 446 Pham T, Claudepierre P, Deprez X, et al. Anti-TNF alpha therapy and safety monitoring. Clinical tool guide elaborated by the Club Rhumatismes et Inflammations (CRI), section of the French Society of Rheumatology (Societe Francaise de Rhumatologie, SFR). Joint Bone Spine 2005;72(Suppl 1):S1-S58. 447 GUIPCAR Group. Clinical practice guideline for the management of rheumatoid arthritis in Spain. 2007 Available at: http://www.guideline.gov/content.aspx?id=12995. Accessed November, 2010. 448 GUIPCAR Group. Clinical practice guideline for the management of rheumatoid arthritis. 2001 Available at: http://www.taiss.com/publi/absful/guipcar%20libro.pdf. Accessed November, 2010. 449 van der Bijl AE, Goekoop-Ruiterman YP, de Vries-Bouwstra JK, et al. Infliximab and methotrexate as induction therapy in patients with early rheumatoid arthritis. Arthritis Rheum 2007;56(7):2129-2134. 450 Buch MH, Marzo-Ortega H, Bingham SJ, et al. Long-term treatment of rheumatoid arthritis with tumour necrosis factor alpha blockade: Outcome of ceasing and restarting biologicals. Rheumatology (Oxford) 2004;43(2):243-244. 451 Quinn MA, Conaghan PG, O'Connor PJ, et al. Very early treatment with infliximab in addition to methotrexate in early, poor-prognosis rheumatoid arthritis reduces magnetic resonance imaging evidence of synovitis and damage, with sustained benefit after infliximab withdrawal: Results from a twelve-month randomized, double-blind, placebo- controlled trial. Arthritis Rheum 2005;52(1):27-35. 452 van der Kooij SM, Goekoop-Ruiterman YP, de Vries-Bouwstra JK, et al. Drug-free remission, functioning and radiographic damage after 4 years of response-driven treatment in patients with recent-onset rheumatoid arthritis. Ann Rheum Dis 2009;68(6):914-921. 453 Hyrich KL, Lunt M, Watson KD, et al. Outcomes after switching from one anti-tumor necrosis factor alpha agent to a second anti-tumor necrosis factor alpha agent in patients with rheumatoid arthritis: Results from a Large UK National Cohort Study. Arthritis Rheum 2007;56(1):13-20. 454 Carmona L, Gomez-Reino JJ, on behalf of the BIOBADASER Group. Survival of TNF antagonists in spondylarthritis is better than in rheumatoid arthritis. Data from the Spanish registry BIOBADASER. Arthritis Res Ther 2006;8(3):R72. 455 Weinblatt M, Bathon J, Kremer J, et al. Efficacy and safety of up to 10 years of etanercept therapy in North American patients with early and long-standing rheumatoid arthritis [abstract]. JAAD 2009;60(3 Suppl 1):AB56. 456 Vander Cruyssen B, Durez P, Westhovens R, et al. Seven-year follow-up of infliximab therapy in rheumatoid arthritis patients with severe long-standing refractory disease: Attrition rate and evolution of disease activity. Arthritis Res Ther 2010;12(3):R77. 457 Delabaye I, De Keyser F, REMITRACT study group. 74-week follow-up of safety of infliximab in patients with refractory rheumatoid arthritis. Arthritis Res Ther 2010;12(3):R121.    211 458 Filippini M, Bazzani C, Zingarelli S, et al. Anti-TNF alpha agents in elderly patients with rheumatoid arthritis: A study of a group of 105 over sixty five years old patients. Reumatismo 2008;60(1):41-49. 459 Figueiredo IT, Morel J, Sany J, et al. Maintenance and tolerability of infliximab in a cohort of 152 patients with rheumatoid arthritis. Clin Exp Rheumatol 2008;26(1):18-23. 460 Navarro-Sarabia F, Ruiz-Montesinos D, Hernandez B, et al. DAS-28-based EULAR response and HAQ improvement in rheumatoid arthritis patients switching between TNF antagonists. BMC Musculoskelet Disord 2009;10:91. 461 King M, Nazareth I, Lampe F, et al. Impact of participant and physician intervention preferences on randomized trials: A systematic review. JAMA 2005;293(9):1089-1099. 462 Preference Collaborative Review Group. Patients' preferences within randomised trials: Systematic review and patient level meta-analysis. BMJ 2008;337:a1864. 463 Janevic MR, Janz NK, Dodge JA, et al. The role of choice in health education intervention trials: A review and case study. Soc Sci Med 2003;56(7):1581-1594. 464 McPherson K, Britton A. Preferences and understanding their effects on health. Qual Health Care 2001;10(Suppl 1):i61-i66. 465 Hunot VM, Horne R, Leese MN, et al. A cohort study of adherence to antidepressants in primary care: The influence of antidepressant concerns and treatment preferences. Prim Care Companion J Clin Psychiatry 2007;9(2):91-99. 466 Antivalle M, Mutti A, Marrazza M, et al. Patient preferences and compliance during rheumatoid arthritis treatment with tumor necrosis factor antagonists (THU0373) [Conference Proceedings]. 2005 Available at: http://www.abstracts2view.com/eular/view.php?nu=EULAR05L1_2005THU0373. Accessed December, 2011. 467 Ariza-Ariza R, Navarro-Sarabia F, Hernandez-Cruz B, et al. Dose escalation of the anti- TNF-alpha agents in patients with rheumatoid arthritis. A systematic review. Rheumatology (Oxford) 2007;46(3):529-532. 468 Huang X, Gu NY, Fox KM, et al. Comparison of methods for measuring dose escalation of the subcutaneous TNF antagonists for rheumatoid arthritis patients treated in routine clinical practice. Curr Med Res Opin 2010;26(7):1637-1645. 469 Kristensen LE, Geborek P, Saxne T. Dose escalation of infliximab therapy in arthritis patients is related to diagnosis and concomitant methotrexate treatment: Observational results from the South Swedish Arthritis Treatment Group register. Rheumatology (Oxford) 2009;48(3):243-245. 470 Favalli EG, Marchesoni A, Colombo GL, et al. Pattern of use, economic burden and vial optimization of infliximab for rheumatoid arthritis in Italy. Clin Exp Rheumatol 2008;26(1):45-51. 471 Carter C, Rizzo J, Gunnarsson CL, et al. Rates and economic consequences of rheumatoid arthritis patients receiving less-than-recommended labeled dosing of adalimumab and etanercept [abstarct]. J Manag Care Pharm 2010;16(7):544-545.    212 472 Morgan S, Coombes M. Income-based drug coverage in British Columbia: Towards an understanding of the policy. Healthc policy 2006;2(2):92-108. 473 British Columbia Ministry of Health. Special Authority request; Abatacept/Adalimumab/Certolozumab/Etanercept/Golimumab/Infliximab Rheumatoid Arthritis (Initial or Switch). 2011 Available at: https://www.health.gov.bc.ca/exforms/pharmacare/5345fil.pdf. Accessed May, 2012. 474 British Columbia Ministry of Health. PharmaCare trends 2009/10. 2010 Available at: http://www.health.gov.bc.ca/pharmacare/pdf/PCareTrends2009-10.pdf. Accessed July, 2011. 475 The Canadian Institute for Health Information (CIHI). Drug use among seniors on public drug programs  in Canada, 2002 to 2008 (National prescription drug utilization information system  database). 2010 Available at: https://secure.cihi.ca/estore/productFamily.htm?pf=PFC1404&locale=en&lang=EN&medi atype=0. Accessed January, 2012. 476 Barker C. The mean, median, and confidence intervals of  the Kaplan-Meier survival estimate – Computations and applications. Amer Statistician 2009;63(1):78-80. 477 Zhong M, Hess KR. Mean Survival Time from Right Censored Data (Article 66). 2009 Available at: http://biostats.bepress.com/cobra/ps/art66. Accessed September, 2011. 478 Everitt BS, Pickles A. Chapter 8: Survival Analysis. In: Everitt BS, Pickles A, eds. Statistical Aspects of the Design and Analysis of Clinical Trials (electronic). Rev. ed. London, England: Imperial College Press; 2004:203-249. 479 WorkSafeBC. About us. 2012 Available at: http://www.worksafebc.com/about_us/default.asp. Accessed Febraury, 2012. 480 Iezzoni LI. Assessing quality using administrative data. Ann Intern Med 1997;127(8 Part 2):666-674. 481 Paulus HE, Egger MJ, Ward JR, et al. Analysis of improvement in individual rheumatoid arthritis patients treated with disease-modifying antirheumatic drugs, based on the findings in patients treated with placebo. The Cooperative Systematic Studies of Rheumatic Diseases Group. Arthritis Rheum 1990;33(4):477-484. 482 Bo H. Sequential analysis of clustered survival data by marginal methods. United States: Rutgers The State University of New Jersey, New Brunswick 2008. 483 Doran MF, Pond GR, Crowson CS, et al. Trends in incidence and mortality in rheumatoid arthritis in Rochester, Minnesota, over a forty-year period. Arthritis Rheum 2002;46(3):625-631. 484 Kaipiainen-Seppanen O, Aho K, Nikkarinen M. Regional differences in the incidence of rheumatoid arthritis in Finland in 1995. Ann Rheum Dis 2001;60(2):128-132. 485 Maini R, St Clair EW, Breedveld F, et al. Infliximab (chimeric anti-tumour necrosis factor alpha monoclonal antibody) versus placebo in rheumatoid arthritis patients receiving concomitant methotrexate: a randomised phase III trial. ATTRACT Study Group. Lancet 1999;354(9194):1932-1939.    213 486 Moreland LW, Schiff MH, Baumgartner SW, et al. Etanercept therapy in rheumatoid arthritis. A randomized, controlled trial. Ann Intern Med 1999;130(6):478-486.    214 APPENDICES    215 APPENDIX A:  CLINICAL AND PHARMACOLOGICAL BACKGROUND (SUPPORTING MATERIAL) This appendix presents in three parts the essential background about the triad being investigated in this thesis: the disease, the drugs and the treatment. The first part is about the disease of RA. RA is a common progressive disease that is costly in terms of impact on quality of life and health care resource use; hence, research is important to reduce these impacts. The second part presents the differences in the pharmacological characteristics of the TNFα antagonist drugs to support the hypothesis that different drugs have different magnitude of effectiveness. The third part describes several special features of TNFα antagonist treatment that are relevant to the thesis. A.1. The Disease: Rheumatoid Arthritis A.1.1 Clinical Manifestation and Diagnosis RA is a common inflammatory systemic disease that causes progressive morbidity and disability, increased mortality and an excessive economic burden to individuals, their families and the health care system. The disease is characterized by inflammation of the lining surfaces of the joints (synovium), heart (pericardium) and lung (pleura) as well as by rheumatoid nodules and vasculitis. The characteristic manifestation is persistent symmetric polyarthritis    216 affecting mainly the small joints of the hands and feet. Diagnosing RA can be challenging because there are no definitive test findings. A set of clinical and laboratory criteria developed by the American College of Rheumatology (ACR) has therefore become the gold standard for diagnosis [294]. At least four of the following criteria are required to confirm the diagnosis of RA: (a) morning stiffness in and around joints lasting more than one hour, (b) arthritis of three or more joint areas involved simultaneously, (c) arthritis of at least one area in a wrist, metacarpal or proximal interphalangeal joint, (d) symmetrical arthritis involving the same joint areas, (e) rheumatoid nodules, (f) positive serum rheumatoid factor, and (g) radiological changes typical of RA in hand and wrist x-rays. RA is also associated with increased risk of cardiovascular disease, infection, osteoporosis and depression, and may shorten life expectancy by six to ten years [295].    A.1.2 Epidemiology and Economic Burden The worldwide prevalence of RA is estimated to be 0.18-1.07% with an incidence of 10-50 new cases per 100,000 population annually [296]. In North America the estimated prevalence is higher at 0.6-1.2% with an annual incidence of 40 per 100,000 [297,298]. In the Canadian    217 province of British Columbia, prevalence was roughly estimated to be between 0.71%63 [299- 301] and 0.90%64 [301,302]. In general, a substantial decline in RA incidence was observed untl 199565 [296]; however, the incidence seems to be on the rise again in recent years [303,304]. RA is more common in women [305] and there is a secular trend toward a more elderly age of onset [296,306,307].  RA is associated with a significant economic burden since it cannot be cured and the ages of most frequent onset, the fourth or fifth decades of life, are relatively early in the life cycle. Estimates of direct and indirect medical costs of the disease vary greatly. This is due at least in part to methodological differences, such as the type of services included in those assessments. Nevertheless, all studies have concluded that costs of RA management are substantial and that a large portion of the overall cost is indirect (e.g., productivity loss, compensation for work disability and impact of mortality) [308].  The direct annual medical costs in Canada were estimated to be $5,100 per patient in the fiscal year May 1999 to May 2000 [309] increasing to $10,300 in the year 2002 [310]. The introduction of infliximab and etanercept to the Canadian market in 2001 was responsible for most of this difference, with a mean annual cost of $4,200                                                   63 In 2000, 22,710 patients were identified with RA [299,300], census British Columbia population (2001) was 3,907,738 [301]; prevalence 0.71% 64  In 2006 37,161 RA patients were identified [302], census British Columbia population was 4,113,487 [301]; prevalence 0.90% 65  A decrease in the incidence rate of RA was observed in Rochester, Minnesota between 1955 and 1994 [483] and in Finland between the years 1980 and 1995 [484]    218 for antirheumatic monoclonal antibodies and immunomodulators [310]. When both direct and indirect costs of treatment are considered, the total annual cost increased from $9,300 in 1999/2000 to $25,000 in 2002. A.1.3 Drug Therapy in Rheumatoid Arthritis The antirheumatic drugs are a heterogeneous group of drugs from various therapeutic classes with a variety of mechanisms of action. Historically, drugs used in treatment of RA are divided into four categories: (a) nonsteroidal anti-inflammatory drugs (NSAIDs), (b) glucocorticosteroids (c) the ‘old’ or traditional antirheumatic drugs and (d) the ‘new’ antirheumatic drugs. The ‘old’ or traditional drugs are short synthetic drugs from a variety of therapeutic classes. The terms disease modifying antirheumatic drugs (DMARDs) or nonbiologic DMARDs have often been used in the literature to identify drugs from this heterogeneous group, regardless of lack of evidence that these drugs actually change the natural history of the disease. The ‘new’ drugs (sometimes termed biologics or biologic DMARDs) are a heterogeneous group of genetically engineered proteins that are either monoclonal antibodies or other immunomodulators. These drugs act to suppress inflammation by inhibiting various cytokines and cell function. We use the term synthetic antirheumatic drugs to describe ‘older’ drugs, including glucocorticosteroids and the term antirheumatic monoclonal antibodies and immunomodulators to describe the ‘biologic drugs’. Table 18 presents the antirheumatic drugs marketed in Canada and approved for treatment of RA.   219 Table 18: Antirheumatic drugs in Canada  ATC, drug  ATC Level 1 ATC Level 3 ATC Level 4 [28] Mechanism of action [28,311] Synthetic antirheumatic drugs A07EC01 Sulfasalazine  Alimentary tract and metabolism Intestinal antinflammatory agents Aminosalicylic acid and similar agents The mechanism of action of sulfasalazine in RA is not completely known. In vitro, sulfasalazine prevent the stimulation of T cells and the activation of nuclear transcription factor kappa B [312]. Additionally it may inhibit prostaglandin synthesis.  L04AD01 Cyclosporine (Ciclosporin) Antineoplastic and immunomodulating agents Immunosuppressant Calcineurin inhibitors Cyclosporine inhibits calcineurin, which is a calcium and calmodulin-dependent serine/threonine phosphatase. Presentation of antigen to T lymphocytes triggers an increase in intracellular calcium levels and activation of calcineurin. Calcineurin dephosphorylate the cytosolic nuclear factor of activated T cells, so it can translocate into the nucleus and stimulate production of cytokines [313]. L04AX01 Azathioprine L04AX03 Methotrexate Antineoplastic and immunomodulating agents Immunosuppressant Other immunosuppressant Azathioprine is a purine analog that is incorporated into replicating DNA and blokes de novo pathways of purine synthesis, an action which is probably specific to lymphocytes [314]. The mechanism of action of methotrexate in the treatment of RA is unknown but may be more anti-inflammatory than immunosuppressive: inhibition of interleukin-1 activity or other inflammatory cytokines (interleukin-6, TNFα) [315] L04AA13 Leflunomide Antineoplastic and immunomodulating agents Immunosuppressant Selective immunosuppressant Leflunomide inhibits dihydroorotate dehydrogenase and tyrosine kinases, and therefore inhibits pyrimidine synthesis. Since activated lymphocytes expend their pyrimidine pool by about eight folds during proliferation, the drug arrests stimulated calls [316,317].    220 ATC, drug  ATC Level 1 ATC Level 3 ATC Level 4 [28] Mechanism of action [28,311] P01BA02  Hydroxychloroquine   Antiparasitic products, insecticide and repellents Antimalarial Quinolines The drug rapidly enters the cell and the lysosomes and elevates the pH within the lysosome. This results in inactivation of acid proteases. The drug interferes with receptor function, inhibits intracellular processing and secretion of proteins, decreases lymphocyte proliferation and interferes with natural killer T-cell activity. It may interrupt activation of synovial fibroblasts [318] . M01CB01 Sodium aurothiomalate  M01CB03 Auranofin  Musculoskeletal system Specific antirheumatic agents Gold preparations Azathioprine is an derivative of mercaptopurine that antagonizes purine metabolism and may inhibit synthesis of DNA, RNA, and proteins. It may also interfere with cellular metabolism and inhibit mitosis. M01CC01   Penicillamine   Musculoskeletal system Specific antirheumatic agents Penicillamine and similar agents The mechanism of action of penicillamine in RA is unknown. It produces no significant depression in absolute levels of serum immunoglobulins and depresses T-cell activity but not B-cell activity in vitro. Antirheumatic monoclonal antibodies and immunomodulators   L04AB01 Etanercept  L04AB02 Infliximab  L04AB04 Adalimumab L04AB05 Certolizumab pegol L04AB06 Golimumab Antineoplastic and immunomodulating agents Immunosuppressants Tumour necrosis factor alpha inhibitors Refer to Section A.2, page 223    221 ATC, drug  ATC Level 1 ATC Level 3 ATC Level 4 [28] Mechanism of action [28,311] L04AC03 Anakinra L04AC07 Tocilizumab Antineoplastic and immunomodulating agents Immunosuppressants Interleukin inhibitors Anakinra is antagonist of the interleukin-1 receptor. Endogenous interleukin-1 is induced by inflammatory stimuli and mediates a variety of immunological responses, including tissue (cartilage and bone) destruction [319] Tocilizumab is a recombinant humanized antihuman interleukin-6 receptor monoclonal antibody. IL-6 is a multifunctional cytokine, produced by a variety of cell types and involved in T-cell activation, induction of acute phase proteins, and stimulation of hematopoiesis. Elevated serum interleukin-6 levels have been reported in RA. Tocilizumab binds specifically to both soluble and membrane-bound interleukin-6 receptors and causes changes in neutrophils and platelet counts [320]. L01XC02 Rituximab Antineoplastic and immunomodulating agents Other neoplastic agents Monoclonal antibodies Rituximab is a genetically engineered chimeric murine/human monoclonal antibody directed against the CD20 antigen found on the surface of normal and malignant B- lymphocytes. CD20 regulates cell cycle initiation; and, possibly, functions as a calcium channel. The mechanism of suppressing RA is not completely understood. Two alternative hypothesis were suggested: (a) elimination of B lymphocytes which are important players in inflammation or (b) the rituximab-B lymphocyte immune complexes bind effector cells and diminish their binding to other inflammation sites [321]by L04AA24 abatacept Antineoplastic and immunomodulating agents Immunosuppressants Selective immunosuppressants  Abatacept is a recombinant fusion protein comprising the extracellular domain of human cytotoxic T-lymphocyte antigen 4 and a fragment of the Fc domain of human IgG1. It acts by competing with CD28 (located on T lymphocytes) for binding to CD80 and CD86 on antigen presenting cells. The drug blocks the interaction between antigen presenting cells and T cells, which activates the T lymphocytes. As a result of the blocked activation, the production of the following immune response mediators is diminished: TNFα , interleukin-2, interferon gamma and granulocyte-macrophage colony-stimulating factor (GM-CSF) [322].     222 Improved understanding of the pathophysiology of RA has led to three key changes in the use of drug therapy [323]. First, early treatment is often recommended since treatment with synthetic antirheumatic drugs during the first five years after diagnosis has been found to minimize progression of radiographic damage [324-327]. Second, combining antirheumatic drugs has been found to be an effective strategy, especially in patients who have failed monotherapy [323]. Evidence from systematic reviews of methotrexate (MTX) monotherapy versus combination therapy [328], and synthetic antirheumatic drug monotherapy versus combination therapy [329] demonstrated superior therapeutic benefit with combination therapy in patients who failed monotherapy. Lastly, drugs targeting cytokines, such as TNFα or interleukin-1, have been developed and approved for marketing and so are newly available treatment options. The antirheumatic monoclonal antibodies and immunomodulators are mainly recommended for RA patients with inadequate response during combination or sequential therapy with synthetic antirheumatic drugs [253,330] .  First line therapy recommended by the ACR Guideline for all patients with RA regardless of disease duration or activity is MTX or leflunomide monotherapy [253,330]. For patients with poor prognostic features, alternative first line drugs include other synthetic antirheumatic drugs such as hydroxychloroquine, minocycline or sulfasalazine. TNFα antagonists (interchangeably [244]) in combination with MTX were recommended as first line therapy for a select group of patients characterized by early RA, high disease activity and features of a poor prognosis. TNFα antagonists were also recommended as second line therapy for two groups of patients: (a) those with an inadequate response to prior MTX and high disease activity and (b) those with an inadequate response to MTX combination therapy or sequential administration of other synthetic antirheumatic drugs.     223 A.2. The Drugs: TNFα Antagonists In this thesis, we explore persistence (duration of treatment) of three drugs from a single therapeutic class: the TNFα antagonists, specifically infliximab, adalimumab and etanercept66. These drugs block the physiological action of the cytokine TNFα, which plays an important role in the pathogenesis of inflammation and RA (Section A.2.2, page 225). Treatment with TNFα antagonists does not cure patients [44], but rather reduces symptoms and possibly induces remission and prevents complications over the long term. TNFα antagonists were demonstrated to slow the radiographic progression of the disease in short-term RCTs [123,331]. These studies used continuous quantitative scale (radiographic scores) that are weaker predictors of clinical outcomes (mortality, costs and function) than patient function [332]. As well, the clinical outcomes are also questionable since some, but not all, long-term real life studies have demonstrated decreased mortality [333], especially from noninfectious causes [334], improved function, manifest as participation in the work force [335,336] and a decrease in hospital stays [337]. Evidence of harm from RCTs and postmarket experience is of major concern in patients treated with the drugs, especially serious and fatal infections [334,338], malignancy [338,339], neurologic and demyelinating events [340] and others. The                                                  66  Two additional TNFα antagonists were recently approved for RA in Canada golimumab (April 2009) and certolizumab (September 2009). These medications were excluded from the analysis of British Columbia data because they were approved after the end of study identification period (31 December 2008).    224 three individual TNFα antagonists have different pharmacological properties, which may result in differences in therapeutic benefit and harm profiles (Section A.2.4, page 229). A.2.1 The Cytokine TNFα The cytokine TNFα primarily acts in the body as a proinflammation mediator by inducing apoptosis or cell death. High concentrations of TNFα released in response to bacterial endotoxins produce a shock-like syndrome, while chronic exposure to low concentration can cause a wasting syndrome in patients with cancer and other diseases [341-343]. Conversely, TNFα also has a role in regeneration and expansion of central or peripheral nerve cells [344,345] and skeletal muscles [346].  Mainly expressed by macrophages, TNFα is also released by a variety of other cells of the immune system, such as monocytes, neutrophils and lymphocytes in response to bacterial products. It is also expressed by a variety of nonimmune cell, including mast cells, endothelial cells, fibroblasts and neuronal tissue [347]. The TNFα cytokine is primarily produced as a transmembranous protein (homotrimer67) that was historically considered a precursor inactive form of the cytokine [348]. The transmembranous TNFα commonly undergoes proteolysis to                                                  67  Homotrimer is a molecular structure, normally a protein, which consists of three identical subunits. The subunits are bonded to each other by a covalent bond (a chemical bond where the atoms are connected by sharing of at least two electrons).    225 its soluble homotrimeric form by the enzyme TACE (TNFα converting enzyme). The soluble TNFα is considered the active form, although the transmembranous TNFα is also involved in the inflammatory response. The soluble TNFα acts on sites remote from the producing cells, while transmembranous TNFα requires cell-to-cell contact [349]. Both forms of TNFα (transmembranous and soluble) bind to one of two membranous receptors: TNFα receptor type 1 (TNF-R1, p55) which is expressed in most tissues and TNFα receptor type 2 (TNF-R2, p75) which is expressed in the cells of the immune system. The receptor activation process is known to involve ligand-dependent trimerization of the receptor but is not fully understood. Activation of TNF-R1 leads to apoptosis. Activation of both receptors induces indirect recruiting of the TNFα receptor-associated factor (TRAF) family and results in activation of nuclear factor kappa B (NF-kB) transcription factors [341-343]. The protein NF-kB controls gene expression in animal cells, and is involved in immune response to inflammation and infection [350].  A.2.2 TNFα in Rheumatoid Arthritis  TNFα is now considered an important mediator of many autoimmune and inflammatory diseases, including RA. The possible role of this cytokine in the pathogenesis of RA was first suggested in the 1980s, when it was demonstrated to have the potential to degrade cartilage and bone in vitro [351,352].  In addition, RA synovial mononuclear cell cultures produce TNFα and other cytokines [353]. In vivo studies have successfully detected TNFα in the synovial fluid and serum of patients with RA but not in healthy volunteers or patients with osteoarthritis    226 [354-357]. Measured TNFα levels in the synovial fluid of the knee joints of RA patients were found to be higher in patients with bone erosions compared to those without erosions [357]. Additionally, a positive correlation between levels of TNFα and rheumatoid factor has been identified [355]. Finally, polymorphism in TNFα/b gene was shown to be an independent marker of RA susceptibility [358] and polymorphism in TNFα promoter -308 A/G was associated with susceptibility to develop severe RA in Latin American population  [359,360]. Activated macrophages in inflamed synovial membranes tissue secrete TNFα. The TNFα cytokine has multiple roles in the pathogenesis of RA [361] including: (a) release of proinflammatory cytokines, such as interleukins 1, 6 and 23, as well as Granulocyte- Macrophage Colony-Stimulating Factor (GM-CSF); (b) leukocyte accumulation via release of chemokines, activation of endothelial cells and induction and/or maintenance of human leukocyte antigen (HLA) class II expression; (c) induction of angiogenesis68; (d) activation of chondrocytes69 and osteoclasts70 leading to cartilage and bone resorption; (e) prostaglandin E2 production and (f) induction of acute phase response by induction of hepcidin. The emerging knowledge on the physiologic role of TNFα and other cytokines and their importance in the pathogenesis of inflammatory diseases is potentially useful in developing therapeutic targets.                                                  68  Angiogenesis is the process of forming new blood vessels from the existing vessels. 69  Chondrocyte is  a cartilage cell, 70  Osteoclast is a large macrophage that erodes bone matrix and plays an important role in the remodeling of bone and the removal of dead bone [3].    227 A.2.3 From Drug Administration to Clinical Response   Administration of TNFα antagonist drugs produces a cascade of processes and events that lead to clinical response. A conceptual model developed by Nestorove 2005 uses four components to describe this sequence [362] (Figure 17).   Figure 17: Cascade of processes from administration of TNFα antagonists to clinical response   Adapted from Nestorov 2005 [360]   The TNFα kinetics process           Generation Expression by cells Secretion locally/ systematically         Mechanisms of action (subcellular, cellular, and higher levels) The clinical response TNFα antagonists’ pharmacokinetics Absorption (etanercept and adalimumab)    Circulation (infliximab, adalimumab and                          etanercept)    Penetration (“site of action”)  The drug binds to the soluble TNF and to membranous TNF  TACE    228 1. TNFα kinetics consists of TNFα production in response to bacterial products and its expression by immune (and other) cells. The transmembranous TNFα may undergo proteolysis by TACE that results in secretion of soluble TNFα, both systemically and in the tissues/fluids. 2. TNFα antagonist pharmacokinetics is represented by absorption of etanercept and adalimumab (administered subcutaneously) from the administration site to the circulation, probably via the lymphatic system [363]. The drugs are pulled by the flow of interstitial fluid into the highly permeable lymphatic capillaries; hence, time to reach maximum drug concentration in the blood (Tmax) may take several days. Infliximab, on the other hand, is administered intravenously. From the blood, the three TNFα antagonists (infliximab, adalimumab and etanercept) penetrate various tissues and fluids. 3. Binding of the drug to the both soluble and transmembranous TNFα interferes with the existing equilibrium of the cytokine and leads to TNFα redistribution. The total TNFα level (both free and bound) increases [364], whereas the free (biologically active) TNFα decreases [362,364]. Certain concentrations of infliximab [365] and etanercept [366] reduce the rate of TNFα clearance and may prolong TNFα availability and further increase total TNFα level. Nestorove 2005 also suggests that the change in TNFα kinetics influences the pharmacokinetics of drugs [362], but this concept was not presented in details or nor supported by research. 4. Pharmacologic mechanisms of action at subcellular, cellular, and higher levels are triggered by the binding to soluble and transmembranous TNFα and result in clinical    229 response. The mechanisms of action of infliximab, adalimumab and etanercept are presented in Section A.2.4.3, page 233. Knowledge of the pharmacokinetic properties of the TNFα antagonists is important in understanding differences in therapeutic benefit and harm between the drugs because absorption and distribution of drugs precede both pharmacodynamics action and clinical response. A.2.4 Differences in Pharmacological Characteristics of the Individual TNFα Antagonists Although the three individual TNFα antagonists are considered one therapeutic class, the class is diverse in pharmacological and pharmacokinetic characteristics (see Table 19, page 237 and Table 20, page 239). These differences may affect the therapeutic benefit and harm, as described below.  A.2.4.1 Structure  Infliximab and adalimumab are full-length immunoglobulin G type 1 (IgG1) monoclonal antibodies (Figure 18) with identical human derived fragment crystallizable (Fc) region. However, these two drugs have a different antigen-binding site. Infliximab is a chimeric protein, contains a mouse-derived antigen-binding site in addition to the human derived Fc region while adalimumab is fully human (Figure 19). Etanercept, on the other hand, is a fusion protein composed of a dimer of the extracellular portions of human tumour TNF-R2 fused to    230 the Fc region of human IgG1 (Figure 19) [188]. The Fc region of etanercept is different from that of infliximab and adalimumab.  Figure 18: The structure of an antibody      Immunoglobulin is a protein that contains two heavy (long) chains and two light (short) chains. The base is called the Fc region, and is composed of two heavy chains. This region binds to a specific class of receptors and other immune molecules to mediate the physiological effects.  The antigen-binding fragment (Fab) region is the site that can bind an antigen. The two Fab regions are located on each of the two arms of the antigen molecule.  Heavy Chains Light  Chain Fab Fc Antigen Binding Site    231 Figure 19: The structure of the TNFα antagonists  A. Infliximab                B. Adalimumab      C. Etanercept  A.2.4.2 Pharmacokinetics  The steady-state concentration-time profiles of etanercept and adalimumab are smoother than infliximab, which may be a desirable quality (Figure 20) [42,188]. The profiles are a result of the route of administration and the serum half-life of the drugs (presented in Table 20, page 239). The wide fluctuations in serum concentrations of infliximab are due to administration in relatively large intravenous boluses. The maximum steady state concentrations of infliximab 3 mg/kg administered intravenously are about 40 times greater those of etanercept (administered     TNFα homotrimer Mouse variable region Human IgG1       232 subcutaneously 25 mg twice a week) and roughly 13 times greater than whose of adalimumab (administered subcutaneously 40 mg every other week) [42]. Consequently, infliximab is associated with an increased risk of exceeding the maximum tolerated concentration at peak times (and concomitant increased risk for harm) or of reaching suboptimal concentrations at trough times (and impaired therapeutic benefit). In addition, the excessive inhibition of TNFα by infliximab may knock out macrophage function and result in improved therapeutic benefit and/or increased susceptibility to infection [189]. Figure 20: Steady-state concentration-time profiles of infliximab, adalimumab and etanercept   The figure represents the simulated steady-state serum concentrations of infliximab, etanercept and adalimumab based on a model developed using published pharmacokinetic parameters and dosing practice [188].  Reprinted from Pharmacology & Therapeutics 117(2), D. Tracey, L. Klareskog, E. H. Sasso, J. G. Salfeld, P. P. Tak Tumour necrosis factor antagonist mechanisms of action: A comprehensive review, Pages No 244- 79., Copyright (2008), with permission from Elsevier (License Number: 2740470467986)     233 TNFα-Drug Complexes and Clearance TNFα-drug complexes, which are antigen-antibody complexes, are typically cleared by a combination of Fc receptor-dependent mechanisms (in the reticuloendothelial system, spleen and liver) and Fc-dependent mechanisms (intracellular degradation and renal filtration). The clearance rate of etanercept-TNFα complex is probably slower than that of infliximab or adalimumab complexes [188]. In addition, these complexes increase the level of circulating TNFα, in a dosage related manner [367-369]; however, the complexes lack bioactivity. TNFα-etanercept complexes are less stable compared to infliximab complexes [370]; hence, the TNFα may be redistributed from the inflamed tissues to other tissues. A.2.4.3 Mechanism of Action Two biological substances are the main targets for TNFα antagonists: TNFα and lymphotoxin alpha. In general, etanercept is distinctive in its faster binding to TNFα [188], inability to bind the monomer form of soluble TNFα [370] and binding to lymphotoxin alpha. An ideal TNFα antagonist blocks the activity of both forms of TNFα (soluble and transmembranous) as both are involved in inflammation and the pathogenesis of the disease. Each infliximab or adalimumab molecule can bind up to two TNFα molecules (soluble or transmembranous) and a single TNFα homotrimer can bind up to three molecules of infliximab or adalimumab while etanercept binds in a 1:1 ratio [43,188,370,371] (Figure 21).  At high, pathological concentrations of soluble TNFα (higher than 0.2 ng/ml), the three drugs have comparable potency to neutralize TNFα. However, at lower and nonpathological TNFα concentration (around 0.1 ng/ml), etanercept is 20 fold more potent [372]. Studies on binding to transmembranous TNFα are limited by the low level of transmembranous TNFα on    234 normal monocytes/macrophages and T cells, even after activation. This is mainly a result of proteolysis of most of the transmembranous TNFα by the enzyme TACE (Section A.2.1, page 224). Consequently, the observed differences in binding to transmembranous TNFα may reflect differences in the activity of TACE under varying experimental conditions and not real differences between drugs [188].  Etanercept binds to lymphotoxin alpha, with an affinity that is similar or even higher than to soluble TNFα. This may results in unneutralizing the soluble TNFα [188]. The clinical significance of this is not clear; however some researchers proposed that manipulation of the lymphotoxin pathway may result in less severe adverse effects [373], especially granulomatous Figure 21: Binding of TNFα antagonists to TNFα   Each infliximab or adalimumab molecule can bind up to two TNFα (soluble or transmembranous) homotrimers; a single TNFα homotrimer can bind up to three molecules of infliximab or adalimumab. Etanercept binds in a 1:1 ratio. Adapted from Rigby 2007 [43]  Binding of soluble TNFα                   Binding of trans-membranous TNFα Etanercept Etanercept    Infliximab, adalimumab Binding of soluble TNFα           Binding of trans-membranous TNFα TNFα       Infliximab or      adalimumab      235 infection [189]. Lymphotoxin alpha was hypothesized to attenuate the function of TNFα [190]; therefore, blocking both lymphotoxin alpha and TNFα activity could result in higher activity of residual TNFα, compared to blockage of TNFα alone. A.2.4.4 Biological Effects Binding to TNFα produces a triad of biological effects: (a) complement-dependent or antibody- dependent cellular toxicity that leads to cytolysis (cell outburst), (b) apoptosis (cell death) and (c) inhibition of interferon gamma production. There are several differences between the TNFα antagonist drugs with regard to these biologic effects. Complement-dependent cytolysis is a process in which the complement binds to proteins on the cell membrane and consequently pores are formed in the cell membrane. Pores allow water to enter the cell and eventually the cell bursts. Infliximab and adalimumab are capable of complement fixation resulting in complement-dependent cytolysis [43,188,189]. Additionally these two drugs are capable of Fc-receptor binding [188] and induce antibody-dependent cellular toxicity [43,189]. In this process, cells of the immune system act against other cells to which antibodies are bound, and cause break down of the antibody-bound cells. The two cytolysis processes are associated with monocytopenia (decrease in the amount of monocytes, which are part of the immune system) that may increase susceptibility to granulomatous infections, mainly tuberculosis. Etanercept, on the other hand, does not fix complement in vitro [374] and does not bind to the Fc-receptors [40]. Infliximab and adalimumab can cause apoptosis of activated T cells and circulating monocytes [189,375], which is a potentially desirable property in chronic inflammation. The mechanisms    236 suggested to explain this phenomenon include either binding to soluble TNFα and depriving cells of survival signals mediated by TNF-R1 signaling, or cross-linkage with transmembranous TNFα which induces a unique signal [43]. The capability of etanercept to induce apoptosis of human monocytes is questionable, as study results were inconsistent [375,376]. Interferon gamma production is inhibited by infliximab and adalimumab but not etanercept [189]. Interferon gamma is produced by T cells and is required for host defense against tuberculosis. The possible excessive risk for tuberculosis posed by infliximab may be a result of inhibition of both TNFα and interferon gamma. A.2.4.5 Immunogenicity  Protein-based drugs such as TNFα antagonists may induce production of antidrug antibodies (‘acquired drug resistance’). This in turn may reduce clinical efficacy due to inactivation and/or rapid clearance of the drug as well as increase the risk of harm, e.g.,. infusion or injection site reactions [188]. Chimeric proteins, which consist of both human and nonhuman parts, including infliximab, are more immunogenic compared to only-human proteins [39,377]. While increased doses of infliximab are associated with decreased immunogenicity [75], concomitant MTX reduces the immunogenicity of all drugs, probably due to its immunosuppressive effects [39,75,120].    237  Table 19: Pharmacological characteristic of the individual TNFα antagonists Drug characteristics Infliximab [186,207] Adalimumab [186,378] Etanercept [186,379] Structure A chimeric antibody that contains murine derived antigen binding variable region (25%) and human IgG1 constant region (75%)  A recombinant monoclonal antibody that contain two kappa light chain and two IgG1 heavy chain 1330 amino acids A human fusion protein that contains two soluble domains of human TNF-R2 and Fc portion of human IgG1 934 amino acids Weight (killodaltons) 149 148 150 Binding  and affinity Binds to soluble and transmembranous TNFα  Affinity: KD  1.17 to 4.5*10 -10  molar Binds to soluble and transmembranous TNFα  Affinity: KD 7.05 * 10 -11  to 1.0 * 10 -10  molar  Binds to soluble and transmembranous TNFα and lymphotoxin alpha Affinity: KD 2.35*10 -11  to 1.15*10 -9  molar Production Recombinant cell line cultured by continuous perfusion Recombinant DNA techniques in mammalian cell expression system Purified by a process that includes specific viral inactivation and removal steps Recombinant DNA techniques in Chinese hamster ovary mammalian cell expression system Immunogenicity Frequency of anti-infliximab antibodies (also known as  human antichimeric antibodies (HACA)) 8-68%[186,187] The prevalence increases with prolonged persistence on infliximab [380] Lower doses of infliximab is associated with higher frequency of HACA [75,120,380-383] Concomitant MTX is associated with decreased incidence of antibody formation [75] The presence of HACA was associated Frequency of anti-adalimumab antibodies in RA 1-87% [186,187]. After three-years of treatment, 28% of the patients developed anti-adalimumab antibodies [387] Concomitant MTX is probably associated with decreased prevalence of antibody formation [388,389] The effect on efficacy is controversial [389-391], while the effect on adverse events is not clear [391] Incidence of anti-etanercept antibodies in clinical trials of RA 2-6% [186,187] with varying rate resulted from different assay No effect was demonstrated on efficacy or adverse events [392]    238 Drug characteristics Infliximab [186,207] Adalimumab [186,378] Etanercept [186,379] with decreased efficacy [381,383,384]; however, the association with adverse events in RA patients is not well studied [385,386] Elimination Not characterized [393] Forms a number of immune complexes of varying weight with TNFα. A stable molecule (three alternating adalimumab and TNFα in a circular chain) is easily removed from the circulation. The complex etanercept-TNFα is metabolized through peptide and amino acid pathways with either recycling of amino acids or elimination in bile and urine [394] DNA - Deoxyribonucleic acid ; IgG1 - Type 1 human immunoglobulin; KD - the antigen dissociation constant, smaller values reflect greater affinity; MTX-methotrexate; RA- rheumatoid arthritis; TNFα – tumour necrosis factor alpha    239 Table 20: Pharmacokinetics characteristics of individual TNFα antagonists Drug characteristics Infliximab  [207,395,396] Adalimumab [76,378,397,398]  Etanercept [379,394,399-401]   Route of administration Intravenous over two hours Subcutaneous Subcutaneous Recommended dose (adults patients with RA, product monograph) 3 mg/kg body weight, at week 0, 2, 6 and then every eight weeks. For patients with an incomplete response, the dose might be adjusted to 5 mg/kg every eight weeks or 3mg/kg every four weeks. 40 mg administered EOW. For patients with an incomplete response, the dose intensity may be increased to 40 mg every week   50 mg per week. Higher doses are not recommended.   Cmax  RA patients, single dose of three mg/kg , no MTX 78.3±26.4 g/ml (n=14)   RA patients, ,multiple dosing of three mg/kg at week 54 25% < 0.001 g/ml Healthy subjects single dose 40 mg  4.7±1.6 g/ml RA patients, single dose of 25 mg 1.07 g/ml (n=25) RA patients, multiple doses 25 mg BIW 2.4±0.99 g/ml (n=23) Tmax  Not applicable (intravenous administration) Healthy subjects single dose 40 mg  131±56 h RA patients, single dose of 25 mg 69.2±33.8 h (n=25) RA patients, multiple doses 25 mg BIW 32.1±27.3 h (n=23) AUC  RA patients, single dose of three mg/kg, no MTX 11765±4800 g*h /ml (n=14)  RA patients, single intravenous dose of 0.5 mg/kg 2729±707 g*h/ml RA patients, single dose of 25 mg 201.7±94.3 g*h/ml RA patients, multiple doses 25 mg BIW AUC0-72 h 143.6±57.2 g*h/ml Clearance RA patients, single dose of 3 mg/kg, no MTX  18.1±5.8 ml/h (n=14)  RA patients 12 ml/h RA patients: MTX reduced clearance by 29% (single dose) to 44% (multiple dosing) Higher clearance in the presence of antibodies Lower clearance with increasing age RA patients, single dose 25 mg (n=25) 160±80 ml/h    240 Drug characteristics Infliximab  [207,395,396] Adalimumab [76,378,397,398]  Etanercept [379,394,399-401]   Volume of distribution   4.3±2.5 L (5 mg/kg intravenous)  4.7-6.0 L (0.25-10 mg/kg intravenous) 8.0 L  (sum of central and peripheral compartments, 2-25 mg intravenous or subcutaneous, single dose or twice a week) Half-life RA patients, single dose of 3 mg/kg no MTX  218±126 h (n=14)  RA patients, single dose 40 mg mean 336 h (240-480 h in different studies) RA patients, single dose of 25 mg 102.3±30.1 h (n=25) RA patients, multiple doses 25 mg BIW 93.7±18.6 h (n=23) Average absolute bioavailability  Absolute bioavailability (administered intravenously) RA patients, single dose 40 mg, 64%  RA patients, 63% Steady state pharmacokinetics (Figure 20, page 232) Data for the steady-state pharmacokinetic of infliximab are limited [41], no systemic accumulation of the drug was observed after repeated treatment with 3 mg/kg to 10 mg/kg at four or 8-week intervals  When infliximab 3 mg/kg every 8 weeks (in RA patients) achieves steady state, the peak drug concentrations are at least 50-fold greater than the median trough concentrations, Trough concentrations were undetectable in 22% to 30% of patients [402]. Drug efficacy is probably compromised at low level of infliximab [120,402] RA patients, 40 mg EOW: the mean steady state concentrations with MTX 8-9 g/ml and without MTX - 5 g/ml The steady state concentrations are greater than the clinically effective serum concentrations (0.8-1.4 g/ml) [403] Etanercept is absorbed slowly and reaches steady state plasma concentrations after 2-4 weeks [394].  Serum concentration profiles at steady state were comparable among patients with RA on 50 mg once weekly and 25 mg twice weekly. At steady state, the volume of distribution is comparable to with infliximab and adalimumab, which, assuming similar tissue distribution,  implies greater tissue penetration for etanercept  g –microgram; AUC- area under the curve; AUC0-72 h- area under the curve between 0-712 hours after administration; BIW- twice a week; Cmax - peak concentration; EOW- every other week; h- hours; mg- milligram; kg- kilogram; L- liter; MTX- methotrexate; RA- rheumatoid arthritis; Tmax- time to peak concentration   241 A.3. TNFα Antagonists Therapy in Rheumatoid Arthritis  A.3.1  The Efficacy and Safety of the Individual TNFα Antagonists This section consists of two parts. The first part introduces a short summary of knowledge from RCTs on the relative benefit and harm of treatment with TNFα antagonists in RA patients. It focuses on comparisons between drugs rather than  each drug compared to placebo. The second part includes a summary of the main limitations of available knowledge from RCTs of TNFα antagonists in RA patients. The latter justifies the necessity and importance of observational studies conducted using data from routine clinical practice, including research presented as part of this thesis. A.3.1.1 Knowledge Based on Randomized Clinical Trials RCTs are the gold standard research design for establishing causation between drug administration and therapeutic outcomes (benefit and harm). This is a reflection of design features, which ensure comparability of the control and intervention groups in all known and unknown prognostic factors, except the intervention of interest [404]. Most RA RCTs estimate the therapeutic benefit of TNFα antagonists using disease severity scores or response criteria. One of the most commonly used sets of criteria is the ACR response criteria [405]. These were determined by a consensus of clinicians and researchers to allow uniform measurement of improvement in disease activity to make comparison of RCT results possible. The ACR20 ACR50 and ACR70 response criteria require (a) 20, 50 or 70%  reduction in the number of    242 swollen joints, (b) 20, 50, or 70%  reduction in the number of tender joints, and (c) 20,50, or 70% reduction in three of the following measures: physician global assessment of disease, patient global assessment of disease, patient assessment of pain, level of C-reactive protein (CRP) or erythrocyte sedimentation rate (ESR), and degree of disability in Health Assessment Questionnaire (HAQ) score [405]. ACR20 is measurable but of questionable clinical significance while ACR50 or AR70 are considered clinically important [406]. Although there are no head-to-head RCTs comparing two (or more) individual TNFα antagonists, multiple indirect comparisons have been conducted using different methodologies to pool the data (Table 21). Most meta-analyses did not demonstrate any difference in therapeutic benefit (ACR20, ACR50 and ACR70) between the drugs [231,407-410] while two meta-analyses identified some differences in therapeutic benefit. Wiens et al 2010 [411] concluded that in short-term studies (6 months), treatment with etanercept and adalimumab had greater benefit than infliximab, while in long-term studies (1-3 years) treatment with adalimumab was the most efficacious of the three TNFα antagonists71. Ingham et al 2010 [412], in the only meta-analysis that focused on radiologic findings, concluded that infliximab demonstrated twice the effect size of adalimumab. The difference in the mean change from .                                                  71  Estimates of therapeutic benefit were based on proportion of patients achieving the ACR response criteria. With short-term treatment (6 months), etanercept was associated with higher rates of reaching ACR20 and ACR50 and adalimumab of achieving ACR70. Over a long-term treatment course (1-3 years), adalimumab demonstrated the highest rate of achieving the three sets of criteria.     243 Table 21: Indirect comparisons of efficacy and safety of the individual TNFα antagonists in RA 72    Reference Number of studies included 73  Outcome of Interest Conclusions Efficacy    Hochberg 2006 [410] 4 (infliximab-1; adalimumab-2; etanercept-1) ACR20, ACR50 The three TNFα antagonists have similar efficacy when added to methotrexate.  Relative risk for achieving ACR50    infliximab versus etanercept 0.52 (95% CI 0.06-4.55),    adalimumab versus etanercept 038 (0.05-2.86),    adalimumab versus infliximab 0.74 (0.36-2.13) Gartlehner 2006 [407] 14 (infliximab-4; adalimumab-5; etanercept-5) ACR20, ACR50 Adjusted indirect comparisons indicated no significant differences in efficacy between TNFα antagonist drugs Nixon 2007 [408] 10 (infliximab-2; adalimumab-4; etanercept-4) ACR50 The effect of the individual TNFα antagonists was similar.  Odds ratios of ACR50    infliximab versus etanercept 0.98 (95% CI 0.45-1.93);    adalimumab versus etanercept 0.94 (0.50-1.62),    adalimumab versus infliximab 0.96 (0.48-1.90))                                                  72  The table presents only meta-analyses that were reported in full publications. Abstracts, conference proceeding and posters were not presented. 73  For meta-analyses that included additional antirheumatic monoclonal antibodies and immunomodulators, the number of studies included presented in this table indicates studies with the three TNFα antagonists infliximab, adalimumab and etanercept.    244 Reference Number of studies included 73  Outcome of Interest Conclusions Alonso-Ruiz 2008 [409] 13 (infliximab-4; adalimumab-5; etanercept-5) ACR20, ACR50, ACR70 No evidence that the relative effects of individual drugs differed Singh 2009 [231] 16 (infliximab-4; adalimumab-8; etanercept-4) ACR50 The absolute improvement was 22% with infliximab, 28% with adalimumab and 36% with etanercept. Differences were not statistically significant Ingham 2010 [412] 3 (infliximab-1; adalimumab-1; etanercept-1) Standardized mean differences (mTSS MD) Infliximab demonstrated twice the effect size of adalimumab. Etanercept was not compared to either drug due to heterogeneity issues. Wiens 2010 [411] 21 (infliximab-7; adalimumab-8; etanercept-6) ACR20, ACR50, ACR70 In short-term therapy (6 months), etanercept and adalimumab demonstrated higher efficacy results. With long-term treatment (1-3 years), adalimumab was found to be the most effective of the three TNFα antagonists. Devine 2011 [414] 10 (infliximab-2; adalimumab-4; etanercept-4) ACR50 No significant difference between the individual drugs was observed Safety    Alonso-Ruiz 2008 [409] 13 (infliximab-4; adalimumab-5; etanercept-5) WDAEs, SAEs and serious infections In patients treated with infliximab and adalimumab, WDAEs were more frequent than in patients treated with etanercept. Treatment with infliximab was associated with a higher frequency of SAEs and serious infections Singh 2010 [413] 16 (infliximab-4; adalimumab-8; etanercept-4) SAEs The effect of the individual TNFα antagonists was similar. Odds ratios for serious adverse events    adalimumab versus infliximab 0.76 (95% CI 0.52-1.12).,   adalimumab versus etanercept 0.80 (0.54-1.20),    etanercept versus infliximab 0.95 (0.66-1.37)    245 Reference Number of studies included 73  Outcome of Interest Conclusions Serious infection The effect of the individual TNFα antagonists was similar. Odds ratios for serious infection    adalimumab versus infliximab 0.77 (0.47-1.27),     adalimumab versus etanercept 1.06 (0.63-1.76),   etanercept versus infliximab 0.73 (0.46-1.15) Total adverse events The effect of the individual TNFα antagonists was similar. Odds ratios for total adverse events    adalimumab versus infliximab 0.92 (0.73-1.16),    adalimumab versus etanercept 1.05 (0.83-1.33),   etanercept versus infliximab 0.87 (0.70-1.10) WDAEs Infliximab was associated with more WDAEs. Odds ratios for WDAE    adalimumab versus infliximab 0.50 (0.32-0.78),    adalimumab versus etanercept 0.80 (0.51-1.25),   etanercept versus infliximab 0.63 (0.41-0.95) ACR20- American College of rheumatology 20% improvement; ACR50- American College of rheumatology 50% improvement; ACR70- American College of rheumatology 70% improvement; CI-confidence interval; MD- mean change scores from baseline of mTSS between active and control groups; mTSS- modifications of Total Sharp scores; SAE – serious adverse event; TNFα – tumour necrosis factor alpha; WDAE- withdrawal due to adverse event     246 baseline of modifications of total Sharp scores74 between active and control groups was -6.8 (95% confidence interval [CI] -9.3 to -4.3) for infliximab compared to -2.6 (95% CI -3.8 to 1.4) for adalimumab.  With regard to harm, patients on infliximab experienced a higher rate of withdrawals due to adverse events (WDAE) compared to etanercept, however the risk compared to adalimumab varied across meta-analyses [409,413]. Higher risk of serious adverse events and serious infections with infliximab was observed in one meta-analysis [413] but not in another which observed a similar risk [409]. The interpretation and generalizability of these meta-analysis results are limited by the limitations of the individual RCTs (below) A.3.1.2 The Limitation of Randomized Clinical Trials  RCTs in general are characterized by relatively small sample sizes and short durations of treatment. Consequently, they often fail to detect rare or delayed therapeutic effects or small or modest differences in major outcomes. Trial participants are often unrepresentative of all of the patients who actually take the drug, as was shown in several diseases [415-418]. Participants tend to be healthier with less comorbidities [419] and more compliant with treatment [420]                                                  74  The Sharp score is a method used to score radiological changes in patients with RA. The scoring is based on erosions and joint space narrowing in radiographs of the hands, wrists and feet [332].    247 compared to real life patients. Additionally, the elderly, women, and minorities are all less likely to be enrolled in RCTs [416,421-427]. The available RCTs of TNFα antagonists for patients with RA need to be carefully appraised because of the potential for important limitations, such as the aftermentioned, to affect the generalizability as well as the interpretation of the results.  First, the duration of RCTs is considerably shorter than the lifetime of the disease (usually 1-2 years compared to several decades). This is particularly important in RA, a disease characterized by a variation in disease activity and deterioration over time. The effects observed in a short-term trial may not be significant in the long-term and a drug which had no effect in short-term may prevent complications or progression of the disease over the long-term [45,428].  Second, RCTs report benefit and harm separately, while in real life clinical decisions are based on an integration of the positive and negative expected therapeutic effects and their relative importance to a specific patient. Treatment persistence not only reflects the actual and perceived benefit and harm of drugs, but also incorporates patient preference and non- therapeutic effects (such as cost and drug coverage policy) that influence prescribing decisions. Third, RA patients who participate in RCTs designed to assess therapeutic benefit and harm of TNFα antagonists are not representative of actual RA populations [211] nor of patients treated with TNFα antagonists in routine clinical settings [212]. Several studies demonstrated that among patients treated with TNFα antagonists in routine clinical settings, only 5-79% were eligible for RCTs based on inclusion and inclusion criteria [210-212,429]. Only about five percentage of patients with advanced RA treated by a rheumatologist in Tennessee [211] were    248 eligible to participate in ATTRACT (Anti-TNF Trial in Rheumatoid Arthritis with Concomitant Therapy)75 [402]. Less than a third of patients treated with TNFα antagonists in the German Biologics Register RABBIT [212] were eligible to participate in the major TNFα antagonists RCTs76. Ineligible patients had lower baseline disease activity, more comorbidity, and lower functional status. Among RA patients treated in 11 hospitals in the Netherland, 34- 79% fulfilled the inclusion criteria based on baseline disease activity required in the different RCT [210]. Among RA patients treated by a single French physician, 42% met inclusion criteria for RCTs [429]. Patients eligible to participate in RCTs had an improved response compared to ineligible patients [210,212,429]. The relatively wide range observed in the proportion of patients eligible to participate in RCTs probably reflects differences in treatment policy or prescribing decisions, but also the use of different trials (with various eligibility criteria) as a comparison. The above mentioned observations suggest that the current eligibility criteria in RCTs may be too restrictive; hence, study findings do not apply to most patients treated in clinical practice.  Fourth, mortality and long-term disability are clinically significant end-points, which necessitate decades of follow-up. Instead, the main outcome criteria used in RCTs are the ACR                                                  75  Patients were randomized to infliximab versus placebo, concomitant with methotrexate 76  Eligibility were based on the inclusion and exclusion criteria of the following major RCT that led to the approval of the drugs for RA: the infliximab ATTRACT [485]; the etanercept study by Moreland et al 1999 [486] and TEMPO [123]; and the adalimumab ARMADA [388] and the study by van de Putte et al 2004 [390].    249 response criteria (presented in Section A.3.1.1, page 241), which include clinical and laboratory features. ACR response criteria have borderline significant association with the progression of joint destruction during a 48-week trial (p=0.03, [430]). Some studies examined the association of several components of the response criteria on mortality [431-434]. Nevertheless, no study measured the association between the response criteria (ACR50 or ACR70) and long-term mortality or disability; hence, the validity of these criteria as prognostic factors is suspect.  There are additional differences in the outcomes of interest between RCTs and the routine practice settings. In clinical trials, the main interest is change, while in clinical practice the main concern is health status [435]. Due to the life-long duration of chronic disease, treatment is prolonged, and the starting point (and change from the starting point) is less relevant as years pass and other clinical issues (such as complications or comorbidities) arise [435]. In addition, in RCTs the therapeutic effect is commonly presented relative to the placebo comparator, while in routine clinical setting the effect relies only on changes in the clinical presentation in the individual treated patient. In clinical setting patient measures, such as quality of life and function are considered of more relevance than physician and laboratory measures that are collected in RCTs, mainly because functional measures are predictive of mortality [434,436].  Last, the positive effect of TNFα antagonists in RA patients is likely to be overestimated in RCTs compared to effects observed in clinical settings. Both ACR20 and ACR50 response rates [210,212] and HAQ improvement [435] were less pronounced in patients in clinical settings compared to RCTs. There are several explanations for this observation. First, trial participants are not representative of those treated in routine clinical setting: their disease is    250 less severe (lower baseline DAS28, swollen joint count and tender joint count), their disease duration is longer, they experienced larger number of synthetic antirheumatic drugs and antirheumatic monoclonal antibodies and immunomodulators [212] and have a higher baseline HAQ [435]. The difference in estimated therapeutic effect between trial participants and routine practice population should be considered in the light of regression to the mean. The observed change depends on the baseline level of disease activity as well as treatment effect with a greater observed change with increasing disease activity. For example, when analyzing change in HAQ scores from visit to visit in clinical settings, Wolfe et al 2004 [435] showed the higher the HAQ score in the previous visit, the larger the change between the visits. The score change between visits increased by 0.24 (95% CI 0.18 to 0.28) for each one unit increase in baseline HAQ. The second explanation is that in clinical practice, the effect of the TNFα antagonists is added to prior treatments while RCTs represent a flare design in which patients must experience a flare prior to entering the trial, prior treatment is restricted and patients must be on stable synthetic antirheumatic drugs (including corticosteroids) prior to treatment [435]. This means that the change observed in RCTs is from a flare and not from a steady state.  Lastly, participants in RCTs are more compliant with the treatment compared to routinely treated patients [420,437]. Therefore, greater exposure to the drug overestimates the benefit as compared to the clinical setting.  To conclude, studies of real life effectiveness remain important to determine the most useful role of each TNFα antagonist in clinical practice setting. This is due to the absence of real world head-to-head RCTs, and the limited generalizability of existing RCTs comparing TNFα antagonists to placebo. Multiple outcomes were used to estimate the effectiveness of TNFα antagonists in RA patients treated in routine practice settings. Most of them involved    251 estimation of disease activity using disease activity or response criteria similar to those used in RCTs, quality of life assessment and functional evaluation. Other effectiveness studies have evaluated harm including rates of infection, malignancy or neurologic complications. Measures encompassing both benefit and harm have rarely been used. Such measures include total hospitalizations (studies unavailable), total mortality [334,438,438] and treatment persistence. A.3.2 Indications for Discontinuing TNFα Antagonists From Clinical Practice Guidelines Multiple national and international guidelines, consensus statements and position papers on treatment for RA patients were published in the last decade; however, not all discussed and defined scenarios for discontinuing TNFα antagonists (Table 22, page 253). The indications for discontinuation of TNFα antagonists can be divided into three major groups:  1. Lack of benefit/ inefficacy (nonresponders, lack of efficacy or loss of efficacy) can be determined based two dimensions: disease activity and duration of exposure to the drug. Disease activity is commonly estimated using validated scores [439] such as the Disease Activity Scores (DAS)77 [440] or the  European League Against Rheumatism (EULAR)                                                  77  DAS or DAS28 is calculated based on the number of tender joints, the number of swollen joints, the Erythrocyte Sedimentation Rate (ESR) or the level of C-reactive protein (CRP) and the patient global health measured on Visual Analogue Scale (VAS). A score larger than 5.1 means high disease activity whereas a     252 response criteria78 [441]). Response is ascertained on a threshold level of disease activity achieved with treatment or minimal improvement with drug therapy. Patients who do not achieve these thresholds are considered nonresponders. The second dimension is a minimal period of continuous administration of the drug to ensure sufficient exposure to the drug. Unfortunately, differences in the criteria used to define both dimensions exist, and some of the guidelines define the dimensions vaguely (Table 22, page 253).  2. Harm (adverse events, adverse effects or intolerability) is considered an indication for discontinuation of the drug [238,239,442-444]. Only the minority of the guidelines discuss specific situations. 3. Remission is rarely discussed as an indication for discontinuation and even then is only considered among other treatment options, such as tapering off glucocorticosteroids, decreasing or discontinuing synthetic antirheumatic drugs or decreasing the dose of the TNFα antagonists.                                                                                                                                                            score lower than 3.2 indicates low disease activity. Remission is defined as a DAS28 score lower than 2.6 [430,440]. 78  EULAR response criteria are based on improvement from baseline of DAS28 and level of disease activity based on DAS28 during the follow up. For example, Good response is defined as improvement of more than 1.2 in DAS28 compared to baseline and a DAS28 follow up level of 2.4 or less. Non responders are defined based on either improvement of 0.6 or less in the score, or improvement of 1.2 or less and DAS28 of more than 3.7 at follow up [441].    253 Table 22: Summary of clinical guidelines - indications for discontinuing TNFα antagonists Reference Inefficacy Safety Remission Other The Canadian Agency for Drugs and Technologies in Health [249,250] Disease activity: not defined Duration of treatment: 14-16 weeks Not mentioned “There was insufficient evidence for the panel to provide a recommendation on discontinuation of TNFα inhibitors in patients achieving remission”  The Canadian Rheumatology Association [251,252,445] “discontinued if meaningful improvement is not achieved’ Not mentioned Not mentioned  The American College of Rheumatology [253,330] Not mentioned Not mentioned Not mentioned  The National Institute for Health and Clinical Excellence[442] Disease activity: response was defined as an improvement in DAS281.2 Duration of treatment: six months Not specified Not mentioned  The French Society for Rheumatology [238] Disease activity: response was defined as either DAS28 <3 or DAS28 <5.1 with a an improvement 1.2 Duration of treatment: 12 weeks Compromised safety was mentioned as an indication for discontinue the drug, and specific indications were further discussed by Club Rhumatismes et Inflammation [446] The publication stated that “when a prolonged remission is obtained, a reduction in DMARDs and biotherapies can be considered.” nevertheless, neither the definition of prolonged remission nor an optimal reduction schedule was agreed upon.     254 Reference Inefficacy Safety Remission Other Spanish Clinical Practice Guideline [239,447,448] Disease activity:  response was defined as DAS28<3.2 Duration of treatment: three months  (Alternative therapeutic options are discussed as well) Tumour , optic neuritis, cytopenia and active tuberculosis were mentioned as indications for discontinuation Toxicity was given as an indication for switching to abatacept or rituximab TNFα antagonist discontinuation was mentioned as a treatment option in patients in remission (defined as DAS28<2.6 or SDAI <5) Lactation Recommendations in Italy [443] Disease activity: no response was defined as either DAS >3.7 and DAS improvement ≤ 1.2; or DAS improvement < 0.6 Duration of treatment: 12 weeks  Discontinuation is recommended when serious infections and/or opportunistic infections occur, (septic arthritis, infected prostheses, acute abscess, osteomyelitis, sepsis, systemic fungal infections). Treatment should only be reinitiated if the infections have been treated adequately. Not mentioned  International consensus statement [444] Disease activity: response was defined vaguely as “individually important responses, including patient oriented measures (for example, HAQ-DI, patients global VAS) or physical measures (for example, joint tenderness)” Duration of treatment: 8-12 weeks The following are indications for discontinuation: Clinical evidence of a lupus-like Syndrome Serious infections, including septic arthritis, infected prostheses, acute abscess, osteomyelitis, sepsis, systemic fungal infections, Listeria etc. Treatment should only be reinitiated if the infections have been treated adequately. A demyelinating-like disorder or optic neuritis Pancytopenia and/or aplastic anemia Not mentioned Pregnancy DAS28- Disease Activity Score in 28 joints; DMARDs- disease modifying antirheumatic drugs ; EULAR- the European league against rheumatism; HAQ-DI- Health Assessment Questionnaire Disability Index; SDAI- Simplified Disease Activity Index; TNFα- tumour necrosis factor alpha; VAS- (pain) visual analog scale    255 Published guidelines and consensus statements imply that persistence with TNFα antagonists in RA patients is affected mainly by the benefit and harm of the drug, because benefit and harm are the main considerations in drug discontinuation. Different and sometime vague definitions of inefficacy and harm are expected to lead to differences in prescribing habits of physicians. Different physicians in the same clinical situation may reach a different decision about discontinuing the TNFα.   A.3.3 Reasons for Discontinuing TNFα Antagonists The two main reasons for discontinuing TNFα antagonists are inefficacy (decreased benefit) and harm. The term ‘inefficacy’ is often used broadly to include both primary inefficacy (lack of efficacy, patients who did not respond to the drug) and secondary inefficacy (loss of efficacy in patients who previously responded to the treatment). In a pooled analysis of studies reporting the reasons for discontinuation in patients treated with the three drugs, we found that 86% of patients discontinued TNFα antagonists due to either inefficacy or harm. The proportions of patients treated with infliximab, adalimumab or etanercept were 79%, 88% or 77%, respectively (Table 23, Figure 22 and Table 24). Only a small number of all discontinuations (<2%) were due to remission [151,179]. High rates of  flares and drug reinitiation were observed in these patients: 0% in six months [242], 13-100% after one year of follow-up [240,241,449,450] and 22-71% after two or more years of follow-up [451,452]. Discontinuation due to remission and subsequent reflaring could cause a temporary interruption in treatment persistence followed by reinitiation of the same drug.    256 Persistence with the TNFα antagonists is likely a valid measure of benfit-harm balance in RA patients because nearly all patients who discontinue these drugs do so as a result of inefficacy or harm and only a small proportion (<2%) discontinue due to remission.  Table 23: Reasons for discontinuing TNFα antagonists Reference Drugs Total number of patients treated Total number of patients discontinue 79  Percentage who discontinue due to inefficacy from all discontinuers 80   Percentage who discontinue due to adverse events (safety)   Hyrich 2007 [453] All 6739 2360 36% 43% INF 3037 1273 36% 41% ADA 876 265 41% 37% ETA 2826 405 34% 49% Carmona 2006 [454] INF, ADA, ETA 4006 1095 36% 49%  Hetland 2010 [110] INF, ADA, ETA 2326 1089 67% 30% Du Pan 2009[170] All 2364 courses 803 (653) courses 50% 49% INF 595 (209) 43% 52% ADA 887 (213) 53% 43% ETA 882 (237) 53% 50% Kievit 2011 [156] INF, ADA, ETA 1560 694 41% 35%                                                  79  In parenthesis is the number of patients for whom reason was reported (if different from the total number of patients who discontinue. 80  Includes both primary (no response) and secondary (loss of efficacy) inefficacy.    257 Reference Drugs Total number of patients treated Total number of patients discontinue 79  Percentage who discontinue due to inefficacy from all discontinuers 80   Percentage who discontinue due to adverse events (safety)   Filippini 2010 [172] and Marchesoni 2009 [151] All 1064 659 (405) 44% 48% INF 519 226 46% 47% ADA 303 111 41% 54% ETA 242 68 45% 41% Weinblatt 2009 [455] ETA 772 573 24% 29% Vander Cruyssen 2010 [456] INF 507 298 35% 41% Heiberg 2008 [115] INF, ADA, ETA 847 293 39% 39% Agarwal 2008 [79] INF, ADA, ETA 503 210 (63) 60% 40% Ornbjerg 2010 [179] INF, ADA, ETA 522 210 46% 37% Vander Cruyssen  2006 [181] INF 479 196 33% 47% Delabaye 2010 [457] INF 504 158 26% 47% Brocq 2007[106] All 441 147 59% (NR 27%, LoE 32%) 41% INF 113 69  71% (NR 20%, LoE 51%) 29% ADA 88 23  61% (NR 48%, LoE 13%) 39% ETA 210 55  42% (NR 27%, LoE 15%) 58% Genevay 2007  [116] INF, ADA, ETA 1561 128 58% 33%    258 Reference Drugs Total number of patients treated Total number of patients discontinue 79  Percentage who discontinue due to inefficacy from all discontinuers 80   Percentage who discontinue due to adverse events (safety)   Filippini 2008 [458] INF, ADA, ETA 295 114 37% 53% Fernandez-Nebro 2007 [161] INF, ADA, ETA 161 88 Not reported 58% Mattey 2010 [152] INF, ADA, ETA 162 79 53% 39% Flendrie 2003 81  [108] All 191 63 40% 59% INF 83 35 37% 71% ADA 94 24 46% 46% ETA 14 4 25% 25% Braun-Moscovici  2008 [105] INF 108 43  70% 30% Figueiredo  2008 [459] INF 152 40 35% 53% Bertoli 2008 [68] INF 77 33  12% 41% Voulgari  2005 [155] INF 84 25 36% 60% Ducoulombier 2007 [69] INF 50 15 47% 47% Wendling 2005 [180] INF 14 41 50% (NR 7%, LoE 43%) 29% Conti 2009 [112] INF, ADA, ETA 37 (second course) 12 92% (NR 59%, LoE 33%) 8% ADA- adalimumab; ETA- etanercept; INF- infliximab; LoE- loss of efficacy in patients who responded; ; NR- no response;                                                  81  Some patients discontinued due to more than one reason.    259 Figure 22: Forest plot for the risk of discontinuing TNFα antagonists due to inefficacy and safety (adverse event)     260  Table 24: Discontinuing TNFα antagonists by reason 82    Drug Number  of studies   Number of discontinuations Percentage from all discontinuations Risk ratio (95% CI) for discontinuing due to inefficacy versus discontinuing due to adverse events   Inefficacy Adverse  events Inefficacy and Adverse  events All 14 7407 44% 42% 86% 1.14 (0.92-1.41) Infliximab 13 2503 37% 42% 79% 0.89 (0.73-1.09) Adalimumab 5 636 46% 42% 88% 1.06 (0.86-1.32) Etanercept 6 1759 34% 43% 77% 0.85 (0.69-1.05) CI- confidence interval; n – number of patients                                                  82  Based on Table 23 and Figure 22.   261 A.3.4 Published Comparisons of Drug Persistence Studies comparing persistence with the available TNFα antagonists in RA patients showed discrepant results. While some published studies reported comparable persistence on the three individual drugs [108,113,460], others demonstrated better persistence on infliximab [62,104] or etanercept [110,150,151,161]. To summarize the results and explore possible causes of heterogeneity, we performed a systematic review and meta-analysis of hazard ratios of discontinuing TNFα antagonists in RA patients (unpublished data). The main findings were: (a) We identified 51 studies in which RA patients were included, patients were treated with at least two TNFα antagonists and drug persistence was reported, but seven studies which report pairwise comparison were included in the main meta-analysis (b) substantial heterogeneity was observed in all comparisons; infliximab and adalimumab I-square quantity of 95% (95% CI 91- 97%) (unpublished data), infliximab and etanercept 79% (95% CI 31-93%) (b) . (c) In the research published to date, heterogeneity arises mainly from the use of different types of data- sources and could be caused by the use of different methodological approaches to ascertain discontinuation from disparate sources.  A.3.5 Patient Characteristics Predict Persistence with TNFα Antagonists in Rheumatoid Arthritis The effect of RA patient characteristics on persistence with TNFα antagonists has been explored in a number of studies. In a systematic review and meta-analyses of the effect of patient characteristics (unpublished data), we demonstrate the following:    262 1. Concomitant use of MTX was significantly associated with improved persistence, with pooled hazard ratio concomitant use compared to non-use 0.73 (95% CI 0.56-0.98, four studies, 4742 patients) and pooled hazard ratio for concomitant use compared to monotherapy 0.79 (95% CI 0.65-0.95, three studies and six drug subgroups, 12593 patients).  2. Exposure to an increase in the number of synthetic antirheumatic drugs was associated with statistically significant improved persistence with hazard ratio for one drug increase of 1.12 (95% CI 1.04-1.20, four studies, 5867 patients). We demonstrate insignificant association between persistence and sex (seven studies, 4,925 patients), age (continuous variable, nine studies, 8,371 patients), disease duration (continuous variable, four studies, 2,547 patients), baseline DAS28 score (continuous variable, four studies, 4,097 patients) and baseline HAQ score (continuous variable, four studies, 4,249 patients).  A.3.6 Diversity in