1. Use of Propensity Scores to Assess Comparability of Treatment Groups in a Registry Program 1 Kenneth J Rothman On behalf of Menno V Huisman, Gregory Y.H. Lip, Hans-Christoph Diener, Sergio J Dubner, Chang-Sheng Ma, Kristina Zint, Christine Teutsch, Nils Schoof, Miney Paquette, Eva Kleine, Dorothee B Bartels, Jonathan Halperin and the GLORIA AF study team

2. Disclosure Information Kenneth J Rothman is an employee of RTI Health Solutions, an independent non-profit research organization that does work for government agencies and pharmaceutical companies. GLORIA-AF is sponsored by Boehringer Ingelheim 2

3. GLORIA-AF 3 GLORIA-AF is a large, international, observational registry program of patients with newly diagnosed AF at risk of stroke Study objectives: 1.Characterize patients in various regions of the world newly diagnosed with non- valvular AF at risk for stroke 2.Describe current patterns of antithrombotic treatment 3.Assess effectiveness and safety of antithrombotic treatments

4. Design of GLORIA-AF 4 Patients on dabigatran etexilate Baseline Visit Phase I Cross-sectional All patients Phase II Cross-sectional, longitudinal follow- up of dabigatran patients Phase III Cross-sectional, longitudinal follow-up of all patients Baseline Visit 3M 6M 1YR 2YR 6M 1YR 2YR 3YR Status: Start after comparability assessment of treatment groups in Phase II Status: Currently ongoing Status: Ended Jan 2013

5. Methods in GLORIA-AF In Phase II and III Propensity Score (PS) techniques will be used to assess comparability of the treatment groups (dabigatran compared with VKA). Specifically, we measure the proportion of patients within the range of overlapping PS. 5 Figure: Sebastian Schneeweiss, Jerry Avorn A review of uses of health care utilization databases for epidemiologic research on therapeutics Journal of Clinical Epidemiology Volume 58, Issue 4 2005 323 - 337

6. Methods in GLORIA-AF The decision to begin Phase III will be primarily based on the degree of overlap of the PS. If 95% or more of patients are in the region of overlapping PS, we will proceed to Phase III for that region. Assessment of comparability is made independently in geographic regions. The first assessment occurred in North America. 6

7. Results From November, 2011 through April, 2013, 1672 eligible patients were enrolled in North America. DE was initiated in 536 and VKA in 488 patients. 15% of patients were prescribed other Novel Oral Anticoagulants (NOACs), and aspirin in 11%. 11% of patients did not receive antithrombotic therapy to prevent ischemic stroke.

8. Results Missing/unknown data were imputed using Naïve Bayes classifier. DE, dabigatran etexilate; VKA, vitamin K antagonist. DE (N = 536)VKA (N = 488) n (%) Age class < 65 years132 (24.6)107 (21.9) 65 to < 75 years204 (38.1)144 (29.5) ≥ 75 years200 (37.3)237 (48.6) Gender Male316 (59.0)262 (53.7) Female220 (41.0)226 (46.3) Hypertension Yes, uncontrolled43 (8.0)48 (9.8) Yes, controlled381 (71.1)355 (72.7) No112 (20.9)85 (17.4)

9. Results Missing/unknown data were imputed using Naïve Bayes classifier. DE, dabigatran etexilate; VKA, vitamin K antagonist. DE (N = 536)VKA (N = 488) n (%) Diabetes mellitus No411 (76.7)328 (67.2) Yes125 (23.3)160 (32.8) CHADS 2 Score Low (score=0)39 (7.3)23 (4.7) Moderate (score=1)206 (38.4)133 (27.3) High (score>=2)291 (54.3)332 (68.0)

10. Results – Propensity score distribution by treatment group 10

11. Results The proportion of patients in the overlapping region of the PS between the two treatment groups was 99% when the model contained a pre-specified subset of risk factors for stroke and bleeding. In a sensitivity analysis that included all collected baseline characteristics, the overlapping region comprised 96%. Either result was sufficient to proceed to the comparative Phase III.

12. Conclusions Nearly all patients fell within the overlapping range of propensity scores. Mostly overlapping PS ranges does not guarantee an absence of confounding. It does, however, allow for control of confounding in the data analysis for those factors that are included in the PS model.

13. Back up slides

14. Variables in main PS model Age class Gender Hypertension Diabetes mellitus Previous stroke TIA non-CNS arterial embolism Congestive heart failure MI Presence of complex aortic plaques Peripheral artery disease Abnormal kidney function Hepatic disease Prior bleeding Any drug (HASBLED) Alcohol abuse Coronary artery disease Smoking status Psychosocial factors

15. Variables included in sensitivity analysis Age class Gender Hypertension Diabetes mellitus (Inclusion criteria) Previous stroke Transient ischemic attack (TIA) Non-CNS arterial embolism Congestive heart failure (Inclusion criteria) Myocardial infarction (MI) Presence of complex aortic plaque Peripheral artery disease Abnormal kidney function Hepatic disease Alcohol abuse Coronary artery disease (CAD) Rheumatic Heart Disease Smoking status Psychosocial factors BMI class Hyperlipidemia Hyperthyroidism Implantation of a biological heart valve Left ventricular hypertrophy Pulmonary embolism Other concomitant drugs Antiplatelet drug use Any drug (HAS BLED) Metabolic and anti-Inflammatory therapy Antihypertensive/heart failure and antiarrhythmic therapy Angina pectoris Neurologic disease Cancer Prior bleeding AF ablation Cardioversion Pacemaker at any timepoint HAS BLED risk score class Physician specialty Type of site LAA occlusion or amputation Weight class Race Categorization of AF Type of AF Creatinine clearance class CHADS2 score class CHA2DS2-VASc score class Medical treatment reimbursed by Deep vein thrombosis (DVT) 15