Skip to main content
PMC home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2020 Apr 1.
Published in final edited form as: Clin Pharmacol Ther. 2018 Nov 11;105(4):979–993. doi: 10.1002/cpt.1256

Claims data studies of direct oral anticoagulants can achieve balance in important clinical parameters only observable in electronic health records

Krista F Huybrechts 1, Chandrasekar Gopalakrishnan 1, Jessica M Franklin 1, Kristina Zint 2, Lionel Riou Franca 2, Dorothee B Bartels 2,3, Joan Landon 1, Sebastian Schneeweiss 1
PMCID: PMC6422763  NIHMSID: NIHMS993984  PMID: 30341980

Abstract

Claims databases provide information on the effects of direct oral anticoagulants (DOACs) as used in routine care but may not contain important data on clinical characteristics which may be captured in electronic health records (EHR).

Within a US claims database, we identified patients initiating a DOAC or warfarin between 10/2010–12/2014. 1:1 propensity score (PS) matching was used to balance 78 claims-defined baseline characteristics. We evaluated whether balance was achieved in patient characteristics immeasurable in the claims data study by evaluating the balance in clinical information (using absolute standardized differences (aSD)) from linked EHR data.

From a claims data cohort study of 140,187 patients, 5,935 (4.2%) patients were linked to EHR data. After PS-matching, almost all EHR-defined patient characteristics were well balanced (aSD<0.1). A new user active comparator design with 1:1 PS matching on many patient characteristics improved balance on clinical risk factors observed in EHR but not in claims data.

Keywords: Direct oral anticoagulants, warfarin, administrative data, claims data, linkage, electronic health records, confounding, sensitivity analysis

Introduction

A number of direct oral anticoagulants (DOACs) are being marketed for the prevention of stroke in patients with non-valvular atrial fibrillation (NVAF).1,2 Unlike vitamin K antagonists, DOACs do not require titration towards a narrow therapeutic range.

DOACs were tested for efficacy and safety in large randomized trials in controlled research settings.3. 4. 5. With their widespread use, concerns arose about the representativeness of these trial findings for large patient populations. For example, the time in therapeutic range observed in the warfarin arm and the level of adherence observed in the DOAC arm of the trials may be overly optimistic for many patients in routine care. Large claims data studies were needed in order to fully understand the safety and effectiveness profile of DOACs given their growing use over time.

Pharmacoepidemiological studies based on longitudinal insurance claims data routinely generated in the provision of healthcare for millions of patients have increasingly been utilized to complement randomized controlled trial (RCT) findings6. 7. 8. 9. and provide information on the comparative effectiveness and safety of anticoagulants in routine care settings. This has resulted in a range of claims data studies of varying quality.10. Even high-quality studies that employ the preferred new user active comparator cohort designs with substantial covariate adjustment11. 12. 13. have been criticized for potential confounding by factors not measured in claims data, including underlying bleeding risks, renal function, over-the-counter (OTC) aspirin use, body mass index (BMI), or smoking.14. Such broad opinions which are not empirically substantiated could be refuted if the factors unmeasured in claims data studies were in fact balanced between treatment groups when measured in clinical data repositories, due to study design choices and high-dimensional proxy adjustment.7. 15.

With the wide-spread use of electronic medical records, subsets of patients identified in administrative claims data can be successfully linked to electronic health records (EHR), and the balance of clinical parameters not documented in claims can be assessed across exposure groups. We sought to evaluate the extent to which balance in clinical characteristics unobserved in claims data was achieved in a monitoring program of the safety and effectiveness of DOACs compared to warfarin.

Results

During the study period, we identified a total of 140,187 patients in the claims cohort (26,199 new dabigatran users, 32,595 new rivaroxaban users, 11,322 new apixaban users and 70,071 new warfarin users). From this claims-based cohort we successfully linked 1,130 dabigatran, 1,602 rivaroxaban, 637 apixaban and 2,566 warfarin users leaving a total EHR-linked subset of 5,935 anticoagulant initiators (4.2% of the total claims-based cohort). After 1:1 PS-matching within the EHR-linked subset, there were 846 dabigatran, 874 rivaroxaban, and 355 apixaban initiators (Figure 1). Patients were more often male (62%) and on average almost 70 years of age.

Figure 1.

Figure 1.

Open in a new tab

Flowchart of study population in sequence of exclusions

Claims-defined characteristics in the study population for whom EHR data were available and in patients without available EHR data were well balanced with almost all aSD <0.1, suggesting the EHR-linked subset was representative of the overall study population (Table 1). However, patients in the linked cohort were slightly younger, had a lower prevalence of hemorrhagic stroke, and slightly lower CHADS and CHA2DS2-VASc scores compared to the not-linked cohort. They had a slightly higher number of distinct medications prescribed and number of physician visits. Similarly, high representativeness was found in each of the three linked DOAC cohorts (Table e1).

Table 1:

Selected characteristics of patients successfully linked to EHR data and those not linked

Claims-defined patient characteristics Linked Not Linked Standardized
difference
N=5,935 N=134,252
N/Mean %/SD N/Mean %/SD linked vs. not linked
Age, years (mean, SD) 67.4 11.4 69.5 12.3 −0.17
Age group (N, %)
 18–54 755 12.7% 15,755 11.7% 0.03
 55–64 1,986 33.5% 38,121 28.4% 0.11
 65–74 1,532 25.8% 31,174 23.2% 0.06
 75+ 1,662 28.0% 49,202 36.6% −0.19
Sex (N, %)
 Male 3,647 61.4% 83,215 62.0% −0.01
Comorbidities during baseline (N, %)
 Acute renal disease 364 6.1% 9,084 6.8% −0.03
 Atherosclerosis 1,805 30.4% 40,904 30.5% 0.00
 Cancer 1,190 20.1% 24,416 18.2% 0.05
 Chronic renal insufficiency 574 9.7% 13,217 9.8% −0.01
 Miscellaneous renal insufficiency 14 0.2% 424 0.3% −0.02
 Coronary artery disease (CAD) 1,999 33.7% 45,511 33.9% 0.00
 Deep vein thrombosis (DVT) 304 5.1% 7,382 5.5% −0.02
 Diabetes 1,585 26.7% 33,724 25.1% 0.04
 Diabetic nephropathy 154 2.6% 2,919 2.2% 0.03
 Heart failure (CHF) 977 16.5% 25,043 18.7% −0.06
 Hemorrhagic stroke 1,460 24.6% 39,578 29.5% −0.11
 Hyperlipidemia 3,120 52.6% 68,747 51.2% 0.03
 Hypertension 5,699 96.0% 128,764 95.9% 0.01
 Hypertensive nephropathy 346 5.8% 8,189 6.1% −0.01
 Intracranial bleeding 16 0.3% 259 0.2% 0.02
 Ischemic stroke 442 7.4% 11,001 8.2% −0.03
 Lower/ unspecified GI bleed 238 4.0% 4,515 3.4% 0.03
 Upper GI bleed 36 0.6% 674 0.5% 0.01
 Urogenital bleed 4 0.1% 57 0.0% 0.01
 Other bleeds 245 4.1% 5,082 3.8% 0.02
 Peptic ulcer disease 1,012 17.1% 20,533 15.3% 0.05
 Peripheral vascular disease (PVD) or PVD surgery 242 4.1% 5,800 4.3% −0.01
 Previous TIA 277 4.7% 6,242 4.6% 0.00
 Prior liver disease 277 4.7% 5,182 3.9% 0.04
 Pulmonary embolism (PE) 198 3.3% 4,619 3.4% −0.01
 Recent MI 281 4.7% 6,589 4.9% −0.01
 Old MI 243 4.1% 5,765 4.3% −0.01
 Renal dysfunction 869 14.6% 20,132 15.0% −0.01
 Stroke 512 8.6% 12,835 9.6% −0.03
 Systemic embolism 50 0.8% 1,256 0.9% −0.01
CHADS2 score (mean, SD) 2.0 1.1 2.1 1.1 −0.09
 1 - Low risk (N, %) 2,530 42.6% 50,330 37.5% 0.11
 2 - Intermediate risk (N, %) 1,912 32.2% 45,676 34.0% −0.04
 3 - High risk (N, %) 1,493 25.2% 38,246 28.5% −0.08
CHA2DS2-VASc score (mean, SD) 3.1 1.6 3.3 1.7 −0.11
 1 - Low risk (N, %) 0 0.0% 0 0.0% −
 2 - Intermediate risk (N, %) 1,030 17.4% 20,828 15.5% 0.05
 3 - High risk (N, %) 4,905 82.6% 113,424 84.5% −0.05
HAS-BLED score (mean, SD) 2.4 1.1 2.4 1.1 −0.03
 1 - Low risk (N, %) 1,385 23.3% 29,114 21.7% 0.04
 2 - Intermediate risk (N, %) 2,200 37.1% 49,876 37.2% 0.00
 3 - High risk (N, %) 2,350 39.6% 55,262 41.2% −0.03
Medications during baseline (N, %)
 Aspirin 85 1.4% 1,793 1.3% 0.01
 Aspirin/dipyridamole 26 0.4% 731 0.5% −0.02
 Clopidogrel 649 10.9% 15,727 11.7% −0.02
 Prasugrel 49 0.8% 807 0.6% 0.03
 Ticagrelor 9 0.2% 225 0.2% 0.00
 Other antiplatelet agents 42 0.7% 862 0.6% 0.01
 NSAIDs 1,399 23.6% 28,090 20.9% 0.06
 Heparin 6 0.1% 83 0.1% 0.01
 Low-molecular weight heparins 377 6.4% 9,246 6.9% −0.02
 PGP inhibitors 3,411 57.5% 74,399 55.4% 0.04
 ARB 1,367 23.0% 29,775 22.2% 0.02
 ACE inhibitor 2,072 34.9% 47,245 35.2% −0.01
 Beta blocker 4,250 71.6% 95,643 71.2% 0.01
 Calcium channel blocker 2,415 40.7% 54,623 40.7% 0.00
 Other hypertension drugs 1,611 27.1% 36,247 27.0% 0.00
 Antiarrhythmic drugs (other than amiodarone and dronedarone) 954 16.1% 16,970 12.6% 0.10
 Statin 3,140 52.9% 70,084 52.2% 0.01
 Other lipid-lowering drugs 774 13.0% 16,242 12.1% 0.03
 Diabetes medications 1,446 24.4% 31,074 23.1% 0.03
Healthcare utilization
 Hospitalization in 30 days prior to treatment initiation (N, %) 2,126 35.8% 54,237 40.4% −0.09
 reating prescriber (N, %)
  Cardiologist 1,302 21.9% 25,339 18.9% 0.08
  Primary care physician 1,470 24.8% 33,604 25.0% −0.01
  Other or unknown 3,163 53.3% 75,309 56.1% −0.06
 Number of laboratory tests ordered (mean, SD) 17.2 27.8 14.6 25.6 0.10
 Number of INR (prothrombin) tests ordered (mean, SD) 1.5 4.0 1.4 3.9 0.03
 Number of lipid tests ordered (mean, SD) 0.9 1.4 0.8 1.3 0.11
 Number of creatinine tests ordered (mean, SD) 0.2 0.8 0.2 0.8 −0.02
 Number of medications (mean, SD) 12.8 6.8 11.6 6.4 0.17
 Number of hospitalizations (mean, SD) 0.7 0.8 0.7 0.8 −0.04
 Number of hospital days (mean, SD) 3.7 7.8 3.9 7.7 −0.03
 Number of office visits (mean, SD) 14.6 11.2 12.6 10.3 0.19
 Number of cardiologist visits (mean, SD) 3.7 4.7 3.5 4.8 0.05
 Number of neurologist visits (mean, SD) 0.3 1.3 0.3 1.4 0.03
Open in a new tab
*

Indicators for region, calendar time of cohort entry and non-CV medications all had SDs of <0.1 and were omitted. PS: propensity score; SD: standard deviation; ACE: angiotensin converting enzyme; ARBs: angiotensin receptor blockers.

Even before PS-matching, reasonable balance had been achieved due to the new user active comparator design16. 17. (Table e1). After PS-matching, all claims-based characteristics were very well balanced between DOAC initiators and warfarin initiators in the EHR-linked subset (Table 2).

Table 2.

Selected claims-defined patient characteristics in those successfully linked with EHR and after PS-matching*

Claims-defined patient characteristics Dabigatran Warfarin Standardized
difference		 Rivaroxaban Warfarin Standardized
difference		 Apixaban Warfarin Standardized
difference
N= 846 N= 846 N= 874 N= 874 N= 355 N= 355
N/Mean %/SD N/Mean %/SD N/Mean %/SD N/Mean %/SD N/Mean %/SD N/Mean %/SD
Age, years (mean, SD) 66.8 11.5 66.9 11.4 −0.01 69.2 10.6 68.6 11.4 0.05 68.7 11.3 69.1 10.7 −0.03
Age group (N, %)
 18–54 114 13.5% 117 13.8% −0.01 78 8.9% 93 10.6% −0.06 43 12.1% 33 9.3% 0.09
 55–64 284 33.6% 298 35.2% −0.03 266 30.4% 265 30.3% 0.00 99 27.9% 111 31.3% −0.07
 65–74 218 25.8% 219 25.9% 0.00 248 28.4% 241 27.6% 0.02 95 26.8% 95 26.8% 0.00
 75+ 230 27.2% 212 25.1% 0.05 282 32.3% 275 31.5% 0.02 118 33.2% 116 32.7% 0.01
Sex (N, %)
 Male 519 61.3% 507 59.9% 0.03 527 60.3% 524 60.0% 0.01 198 55.8% 205 57.7% −0.04
Comorbidities during baseline1 (N, %)
 Acute renal disease 35 4.1% 38 4.5% −0.02 59 6.8% 51 5.8% 0.04 19 5.4% 15 4.2% 0.05
 Atherosclerosis 235 27.8% 213 25.2% 0.06 309 35.4% 273 31.2% 0.09 120 33.8% 128 36.1% −0.05
 Cancer 120 14.2% 126 14.9% −0.02 221 25.3% 202 23.1% 0.05 99 27.9% 97 27.3% 0.01
 Chronic renal insufficiency 53 6.3% 45 5.3% 0.04 98 11.2% 90 10.3% 0.03 43 12.1% 35 9.9% 0.07
 Miscellaneous renal insufficiency 1 0.1% 1 0.1% 0.00 2 0.2% 1 0.1% 0.03 1 0.3% 1 0.3% 0.00
 Coronary artery disease (CAD) 261 30.9% 249 29.4% 0.03 338 38.7% 294 33.6% 0.10 128 36.1% 137 38.6% −0.05
 Deep vein thrombosis (DVT) 16 1.9% 19 2.2% −0.02 53 6.1% 45 5.1% 0.04 6 1.7% 7 2.0% −0.02
 Diabetes 197 23.3% 197 23.3% 0.00 265 30.3% 238 27.2% 0.07 116 32.7% 107 30.1% 0.05
 Diabetic nephropathy 8 0.9% 8 0.9% 0.00 26 3.0% 29 3.3% −0.02 11 3.1% 12 3.4% −0.02
 Heart failure (CHF) 134 15.8% 122 14.4% 0.04 162 18.5% 146 16.7% 0.05 58 16.3% 52 14.6% 0.05
 Hemorrhagic stroke 312 36.9% 307 36.3% 0.01 114 13.0% 118 13.5% −0.01 1 0.3% 0 0.0% 0.08
 Hyperlipidemia 390 46.1% 378 44.7% 0.03 492 56.3% 455 52.1% 0.09 213 60.0% 210 59.2% 0.02
 Hypertension 811 95.9% 815 96.3% −0.02 842 96.3% 841 96.2% 0.01 348 98.0% 346 97.5% 0.04
 Hypertensive nephropathy 29 3.4% 22 2.6% 0.05 47 5.4% 41 4.7% 0.03 27 7.6% 19 5.4% 0.09
 Intracranial bleeding 4 0.5% 4 0.5% 0.00 0 0.0% 0 0.0% − 1 0.3% 0 0.0% 0.08
 Ischemic stroke 53 6.3% 56 6.6% −0.01 70 8.0% 69 7.9% 0.00 25 7.0% 23 6.5% 0.02
 Lower/ unspecified GI bleed 27 3.2% 24 2.8% 0.02 38 4.3% 33 3.8% 0.03 16 4.5% 16 4.5% 0.00
 Upper GI bleed 3 0.4% 5 0.6% −0.03 4 0.5% 4 0.5% 0.00 2 0.6% 4 1.1% −0.06
 Urogenital bleed 0 0.0% 0 0.0% − 0 0.0% 0 0.0% − 0 0.0% 0 0.0% −
 Other bleeds 19 2.2% 16 1.9% 0.02 44 5.0% 51 5.8% −0.04 14 3.9% 10 2.8% 0.06
 Peptic ulcer disease 117 13.8% 113 13.4% 0.01 174 19.9% 166 19.0% 0.02 67 18.9% 69 19.4% −0.01
 Peripheral vascular disease (PVD) or PVD surgery 30 3.5% 29 3.4% 0.01 37 4.2% 34 3.9% 0.02 18 5.1% 14 3.9% 0.05
 Previous TIA 32 3.8% 36 4.3% −0.02 41 4.7% 35 4.0% 0.03 12 3.4% 11 3.1% 0.02
 Prior liver disease 24 2.8% 25 3.0% −0.01 49 5.6% 49 5.6% 0.00 18 5.1% 19 5.4% −0.01
 Pulmonary embolism (PE) 11 1.3% 9 1.1% 0.02 30 3.4% 24 2.7% 0.04 2 0.6% 2 0.6% 0.00
 Recent MI 33 3.9% 32 3.8% 0.01 46 5.3% 41 4.7% 0.03 14 3.9% 14 3.9% 0.00
 Old MI 31 3.7% 29 3.4% 0.01 42 4.8% 36 4.1% 0.03 12 3.4% 10 2.8% 0.03
 Renal dysfunction 87 10.3% 75 8.9% 0.05 149 17.0% 134 15.3% 0.05 59 16.6% 51 14.4% 0.06
 Stroke 60 7.1% 63 7.4% −0.01 81 9.3% 79 9.0% 0.01 28 7.9% 29 8.2% −0.01
 Systemic embolism 4 0.5% 4 0.5% 0.00 10 1.1% 9 1.0% 0.01 3 0.8% 3 0.8% 0.00
CHADS2 score (mean, SD) 1.8 1.0 1.8 1.0 0.01 2.1 1.2 2.0 1.1 0.07 2.1 1.2 2.0 1.1 0.07
 1 - Low risk (N, %) 389 46.0% 388 45.9% 0.00 329 37.6% 340 38.9% −0.03 138 38.9% 128 36.1% 0.06
 2 - Intermediate risk (N, %) 288 34.0% 292 34.5% −0.01 276 31.6% 295 33.8% −0.05 113 31.8% 129 36.3% −0.10
 3 - High risk (N, %) 169 20.0% 166 19.6% 0.01 269 30.8% 239 27.3% 0.08 104 29.3% 98 27.6% 0.04
CHADS2−VASc score (mean, SD) 2.9 1.5 2.9 1.5 0.01 3.4 1.7 3.3 1.6 0.09 3.3 1.7 3.2 1.5 0.08
 1 - Low risk (N, %) 0 0.0% 0 0.0% − 0 0.0% 0 0.0% − 0 0.0% 0 0.0% −
 2 - Intermediate risk (N, %) 160 18.9% 162 19.1% −0.01 120 13.7% 125 14.3% −0.02 42 11.8% 43 12.1% −0.01
 3 - High risk (N, %) 686 81.1% 684 80.9% 0.01 754 86.3% 749 85.7% 0.02 313 88.2% 312 87.9% 0.01
HAS-BLED score (mean, SD) 2.2 1.0 2.2 1.0 0.02 2.5 1.1 2.5 1.1 0.04 2.5 1.1 2.4 1.0 0.06
 1 - Low risk (N, %) 237 28.0% 230 27.2% 0.02 166 19.0% 159 18.2% 0.02 69 19.4% 60 16.9% 0.07
 2 - Intermediate risk (N, %) 330 39.0% 346 40.9% −0.04 298 34.1% 334 38.2% −0.09 126 35.5% 157 44.2% −0.18
 3 - High risk (N, %) 279 33.0% 270 31.9% 0.02 410 46.9% 381 43.6% 0.07 160 45.1% 138 38.9% 0.13
Medications during baseline (N, %)
 Aspirin 14 1.7% 12 1.4% 0.02 13 1.5% 14 1.6% −0.01 9 2.5% 8 2.3% 0.02
 Aspirin/dipyridamole 3 0.4% 4 0.5% −0.02 3 0.3% 3 0.3% 0.00 3 0.8% 3 0.8% 0.00
 Clopidogrel 93 11.0% 89 10.5% 0.02 115 13.2% 99 11.3% 0.06 39 11.0% 44 12.4% −0.04
 Prasugrel 6 0.7% 7 0.8% −0.01 7 0.8% 8 0.9% −0.01 6 1.7% 4 1.1% 0.05
 Ticagrelor 1 0.1% 1 0.1% 0.00 0 0.0% 0 0.0% − 0 0.0% 0 0.0% −
 Other antiplatelet agents 4 0.5% 7 0.8% −0.04 7 0.8% 6 0.7% 0.01 3 0.8% 4 1.1% −0.03
 NSAIDs 182 21.5% 179 21.2% 0.01 212 24.3% 215 24.6% −0.01 74 20.8% 80 22.5% −0.04
 Heparin 0 0.0% 0 0.0% − 0 0.0% 0 0.0% − 0 0.0% 0 0.0% −
 Low-molecular weight heparins 17 2.0% 25 3.0% −0.06 24 2.7% 21 2.4% 0.02 4 1.1% 6 1.7% −0.05
 PGP inhibitors 492 58.2% 475 56.1% 0.04 496 56.8% 499 57.1% −0.01 215 60.6% 204 57.5% 0.06
 ARB 188 22.2% 195 23.0% −0.02 221 25.3% 207 23.7% 0.04 90 25.4% 79 22.3% 0.07
 ACE inhibitor 307 36.3% 291 34.4% 0.04 309 35.4% 311 35.6% 0.00 126 35.5% 123 34.6% 0.02
 Beta blocker 623 73.6% 620 73.3% 0.01 627 71.7% 628 71.9% 0.00 256 72.1% 251 70.7% 0.03
 Calcium channel blocker 345 40.8% 347 41.0% 0.00 386 44.2% 389 44.5% −0.01 162 45.6% 157 44.2% 0.03
 Other hypertension drugs 227 26.8% 219 25.9% 0.02 253 28.9% 243 27.8% 0.03 121 34.1% 108 30.4% 0.08
 Antiarrhythmic drugs (other than amiodarone and dronedarone) 136 16.1% 137 16.2% 0.00 114 13.0% 132 15.1% −0.06 54 15.2% 45 12.7% 0.07
 Statin 429 50.7% 430 50.8% 0.00 486 55.6% 463 53.0% 0.05 203 57.2% 194 54.6% 0.05
 Other lipid-lowering drugs 104 12.3% 108 12.8% −0.01 116 13.3% 105 12.0% 0.04 43 12.1% 54 15.2% −0.09
 Diabetes medications 192 22.7% 191 22.6% 0.00 231 26.4% 210 24.0% 0.06 98 27.6% 83 23.4% 0.10
Healthcare utilization
 Hospitalization in 30 days prior to treatment initiation (N, %) 298 35.2% 299 35.3% 0.00 328 37.5% 315 36.0% 0.03 90 25.4% 81 22.8% 0.06
 Treating prescriber (N, %)
  Cardiologist 150 17.7% 149 17.6% 0.00 196 22.4% 193 22.1% 0.01 109 30.7% 119 33.5% −0.06
  Primary care physician 206 24.3% 206 24.3% 0.00 281 32.2% 259 29.6% 0.05 98 27.6% 90 25.4% 0.05
  Other 490 57.9% 491 58.0% 0.00 397 45.4% 422 48.3% −0.06 148 41.7% 146 41.1% 0.01
 Number of laboratory tests ordered (mean, SD) 15.0 22.8 15.8 27.7 −0.03 14.2 19.7 15.0 26.4 −0.03 15.4 32.8 14.2 24.2 0.04
 Number of INR (prothrombin) tests ordered (mean, SD) 0.9 2.7 1.1 2.5 −0.06 0.6 2.2 0.8 2.2 −0.07 0.7 2.7 0.8 1.9 −0.03
 Number of lipid tests ordered (mean, SD) 0.9 1.3 0.9 1.3 −0.05 0.8 1.2 0.8 1.3 −0.01 0.9 1.5 0.8 1.4 0.07
 Number of creatinine tests ordered (mean, SD) 0.2 0.8 0.2 0.8 0.01 0.1 0.6 0.1 0.7 0.00 0.1 0.4 0.1 0.4 0.13
 Number of medications (mean, SD) 12.3 6.7 12.1 6.4 0.03 13.0 6.8 12.8 6.7 0.02 13.2 6.3 13.2 6.9 0.01
 Number of hospitalizations (mean, SD) 0.6 0.7 0.6 0.8 −0.02 0.6 0.7 0.6 0.8 0.01 0.5 0.7 0.5 0.8 0.02
 Number of hospital days (mean, SD) 2.7 4.4 2.7 3.8 −0.01 3.4 6.5 3.4 6.0 0.00 2.6 5.8 2.6 5.8 0.01
 Number of office visits (mean, SD) 13.4 9.9 13.6 10.6 −0.01 14.2 9.9 14.3 10.7 −0.01 14.0 10.2 13.5 9.2 0.05
Open in a new tab
*

Indicators for region, calendar time of cohort entry and non-CV medications all had SDs of <0.1 and were omitted. PS: propensity score; SD: standard deviation; ACE: angiotensin converting enzyme; ARBs: angiotensin receptor blockers.

Most EHR-based clinical patient characteristics that were unobserved in the claims data study were well balanced within each 1:1 PS-matched exposure group (Table 3). Although we observed a relatively large proportion of patients with missing EHR information (ranging from approximately 25% to 98% depending on the variable) (Table 3 and Table e2), the mechanism underlying the missingness appeared to be non-differential between exposure groups for all variables except INR. The large amount of missing information which was differential for INR, a HAS-BLED score component, limits the interpretation of the HAS-BLED score with EHR data.

Table 3.

Selected baseline EHR-based characteristics of study participants after claims-based PS-matching

EHR-observed measurements Dabigatran Warfarin Standardized
difference		 Rivaroxaban Warfarin Standardized
difference		 Apixaban Warfarin Standardized
difference
N=846 N=846 N=874 N=874 N=355 N=355
N/Mean %/SD N/Mean %/SD N/Mean %/SD N/Mean %/SD N/Mean %/SD N/Mean %/SD
Body mass index (BMI; kg/m2)
Missing BMI data (N, %) 205 24.2% 187 22.1% 0.05 121 13.8% 136 15.6% −0.05 36 10.1% 43 12.1% −0.06
Patients with BMI data (N, %) 641 75.8% 659 77.9% −0.05 753 86.2% 738 84.4% 0.05 319 89.9% 312 87.9% 0.06
 BMI (mean, SD) 32.4 8.3 32.5 8.0 −0.01 31.8 7.6 31.9 7.9 −0.02 31.9 7.4 32.4 8.1 −0.06
 BMI category (N, % of pts w/ BMI data)
  Underweight: BMI <18.5 5 0.8% 3 0.5% 0.04 9 1.2% 6 0.8% 0.04 4 1.3% 3 1.0% 0.03
  Healthy weight: BMI 18.5 to <25 89 13.9% 102 15.5% −0.05 117 15.5% 114 15.4% 0.00 46 14.4% 49 15.7% −0.04
  Overweight: BMI 25 to <30 179 27.9% 197 29.9% −0.04 220 29.2% 227 30.8% −0.03 97 30.4% 80 25.6% 0.11
  Obese: BMI >30 368 57.4% 357 54.2% 0.07 407 54.1% 391 53.0% 0.02 172 53.9% 180 57.7% −0.08
   Class 1 obese: BMI 30 to <35 188 29.3% 153 23.2% 0.14 198 26.3% 187 25.3% 0.02 76 23.8% 87 27.9% −0.09
   Class 2 obese: BMI 35 to <40 86 13.4% 99 15.0% −0.05 122 16.2% 108 14.6% 0.04 51 16.0% 48 15.4% 0.02
   Class 3 obese: BMI >40 94 14.7% 105 15.9% −0.04 87 11.6% 96 13.0% −0.04 45 14.1% 45 14.4% −0.01
Smoking
Missing smoking data (N, %) 758 89.6% 749 88.5% 0.03 767 87.8% 771 88.2% −0.01 318 89.6% 315 88.7% 0.03
Patients with smoking data (N, %) 88 10.4% 97 11.5% −0.03 107 12.2% 103 11.8% 0.01 37 10.4% 40 11.3% −0.03
 Never smoked (N, % of pts w/ smoking data) 44 50.0% 48 49.5% 0.01 51 47.7% 48 46.6% 0.02 21 56.8% 16 40.0% 0.34
 Current/past (N, % of pts w/ smoking data) 44 50.0% 49 50.5% −0.01 56 52.3% 55 53.4% −0.02 16 43.2% 24 60.0% −0.34
Alcohol consumption
Missing alcohol consumption data (N, %) 832 98.3% 828 97.9% 0.03 852 97.5% 853 97.6% −0.01 350 98.6% 345 97.2% 0.10
Patients with alcohol consumption data (N, %) 14 1.7% 18 2.1% −0.03 22 2.5% 21 2.4% 0.01 5 1.4% 10 2.8% −0.10
 No consumption (N, % of pts w/ alcohol data) 7 50.0% 11 61.1% −0.23 9 40.9% 13 61.9% −0.43 0 0.0% 5 50.0% −1.41
 Light to moderate consumption2 (N, % of pts w/ alcohol data) 1 7.1% 0 0.0% 0.39 2 9.1% 1 4.8% 0.17 0 0.0% 0 0.0% −
 Heavy consumption3 (N, % of patients w/ alcohol data) 6 42.9% 6 33.3% 0.20 11 50.0% 6 28.6% 0.45 5 100.0% 5 50.0% 1.41
 Consumption of unknown quantify (N, % of pts w/ alcohol data) 0 0.0% 1 5.6% −0.34 0 0.0% 1 4.8% −0.32 0 0.0% 0 0.0% −
Glomerular filtration rate (GFR; ml/min/1.73m2)
Missing GFR data (N, %) 405 47.9% 405 47.9% 0.00 432 49.4% 431 49.3% 0.00 161 45.4% 179 50.4% −0.10
Patients with GFR data (N, %) 441 52.1% 441 52.1% 0.00 442 50.6% 443 50.7% 0.00 194 54.6% 176 49.6% 0.10
 GFR (mean, SD) 85.4 21.9 83.4 23.5 0.09 81.8 24.5 80.0 24.7 0.07 81.1 25.2 79.4 23.9 0.07
 GFR category (N, % of pts w/ GFR data)
  G1: GFR >90 202 45.8% 189 42.9% 0.06 177 40.0% 164 37.0% 0.06 78 40.2% 63 35.8% 0.09
  G2: GFR 60 to 89 175 39.7% 178 40.4% −0.01 175 39.6% 184 41.5% −0.04 76 39.2% 80 45.5% −0.13
  G3a: GFR 45 to 59 43 9.8% 42 9.5% 0.01 47 10.6% 52 11.7% −0.04 21 10.8% 15 8.5% 0.08
  G3b: GFR 30 to 44 19 4.3% 17 3.9% 0.02 35 7.9% 28 6.3% 0.06 13 6.7% 14 8.0% −0.05
  G4: GFR 15 to 29 2 0.5% 14 3.2% −0.20 7 1.6% 11 2.5% −0.06 4 2.1% 2 1.1% 0.07
  G5: GFR <15 0 0.0% 1 0.2% −0.07 1 0.2% 4 0.9% −0.09 2 1.0% 2 1.1% −0.01
Abnormal renal function (N, %) 55 6.5% 68 8.0% −0.06 83 9.5% 88 10.1% −0.02 35 9.9% 35 9.9% 0.00
Duration of atrial fibrillation
 Months (mean, SD) 22.7 35.2 25.4 33.0 −0.08 22.7 34.0 24.0 33.6 −0.04 18.2 28.6 23.1 35.0 −0.15
 <1 year (N, % of pts w/ atrial fibrillation) 224 58.0% 174 49.7% 0.17 206 56.1% 185 55.4% 0.01 113 64.6% 84 57.1% 0.15
 1 to <3 years 78 20.2% 85 24.3% −0.10 74 20.2% 65 19.5% 0.02 26 14.9% 31 21.1% −0.16
 3 to <5 years 43 11.1% 45 12.9% −0.05 49 13.4% 34 10.2% 0.10 21 12.0% 13 8.8% 0.10
 5+ years 41 10.6% 46 13.1% −0.08 38 10.4% 50 15.0% −0.14 15 8.6% 19 12.9% −0.14
Duration of stroke
 Months (mean, SD) 34.6 38.6 34.4 34.0 0.00 32.3 34.9 31.3 32.1 0.03 32.7 40.6 36.2 41.5 −0.09
 <1 year (N, % of pts w/ stroke) 16 34.0% 19 35.8% −0.04 27 39.1% 19 33.3% 0.12 12 46.2% 9 45.0% 0.02
 1 to <3 years 17 36.2% 16 30.2% 0.13 19 27.5% 22 38.6% −0.24 7 26.9% 4 20.0% 0.16
 3 to <5 years 4 8.5% 7 13.2% −0.15 12 17.4% 7 12.3% 0.14 1 3.8% 2 10.0% −0.24
 5+ years 10 21.3% 11 20.8% 0.01 11 15.9% 9 15.8% 0.00 6 23.1% 5 25.0% −0.05
Use of antiplatelets or NSAIDs, incl. OTC use (N, %) 145 17.1% 158 18.7% −0.04 183 20.9% 184 21.1% 0.00 74 20.8% 75 21.1% −0.01
Bleeding history or predisposition (N, %) 31 3.7% 39 4.6% −0.05 28 3.2% 34 3.9% −0.04 8 2.3% 14 3.9% −0.10
International normalized ratio (INR)
Missing INR data (N, %) 747 88.3% 671 79.3% 0.25 809 92.6% 691 79.1% 0.39 327 92.1% 282 79.4% 0.37
Patients with INR data (N, %) 99 11.7% 175 20.7% −0.25 65 7.4% 183 20.9% −0.39 28 7.9% 73 20.6% −0.37
 INR (mean, SD) 1.3 0.5 1.8 0.7 −0.76 1.4 0.7 1.7 0.7 −0.44 1.2 0.5 1.6 0.8 −0.62
 INR category (N, % of pts w/ INR data)
  <1 15 15.2% 14 8.0% 0.22 10 15.4% 14 7.7% 0.24 5 17.9% 4 5.5% 0.39
  1 to <2 74 74.7% 83 47.4% 0.58 46 70.8% 99 54.1% 0.35 22 78.6% 45 61.6% 0.38
  2 to <3 8 8.1% 68 38.9% −0.78 5 7.7% 62 33.9% −0.68 0 0.0% 20 27.4% −0.87
  >3 2 2.0% 10 5.7% −0.19 4 6.2% 8 4.4% 0.08 1 3.6% 4 5.5% −0.09
HAS-BLED score
 HAS-BLED (mean, SD) 1.5 1.0 1.5 1.0 −0.06 1.6 1.0 1.6 1.0 −0.05 1.6 1.0 1.7 1.0 −0.04
 <1 (N, % of all pts) 127 15.0% 115 13.6% 0.04 118 13.5% 103 11.8% 0.05 41 11.5% 36 10.1% 0.05
 1 to 2 607 71.7% 599 70.8% 0.02 599 68.5% 610 69.8% −0.03 246 69.3% 248 69.9% −0.01
 >2 112 13.2% 132 15.6% −0.07 157 18.0% 161 18.4% −0.01 68 19.2% 71 20.0% −0.02
Open in a new tab
*

PS: propensity score; IQR: interquartile; BMI: body mass index; HbA1c: hemoglobin A1c; eGFR: estimated glomerular filtration rate; HDL: high-density lipoprotein; LDL: low-density lipoprotein; BP: blood pressure; HAS-BLED: Labile INR defined as most recent INR <2 or >3 prior to cohort entry.

Smoking status and BMI categories were well balanced with aSDs mostly below 0.05, although smoking status was largely missing (~90%) in the EHR. As some patient numbers get smaller than 20 per treatment group (e.g., smoking in apixaban cohort), chance variation is likely to have led to slightly higher imbalances but with aSD still smaller than 0.40. Alcohol use was recorded in less than 5% of patients and resulted in non-interpretable findings. Estimated GFR was available for about 50% of patients and mean eGFR was well balanced across all groups. eGFR categories also showed very good balance, with aSDs all <0.10 except for a few categories with small numbers of patients. The mean duration of existing AF, which is difficult to fully assess in claims data due to left censoring, was well balanced (aSD <0.10) for dabigatran and rivaroxaban, with some imbalance (aSD=0.15) observed for the apixaban group. NSAID use including recorded over-the-counter (OTC) use though seemingly less completely captured in EHR versus prescription databases, was very well balanced (aSDs < 0.05). INR values were recorded in less than 20% of patients. INR measurements are more likely to be performed amongst warfarin users than among DOAC users. For those patients for whom it was recorded, there was evidence of some imbalance. The HAS-BLED score for bleeding defined by EHR information was very well balanced (Table 3).

Any residual imbalances in EHR-defined clinical variables could potentially cause confounding bias. The potential confounding bias caused by any residual imbalances in EHR-defined clinical variables was negligible for plausible scenarios of unmeasured confounder-outcome (RRCD) associations (Figure 2). For example, if current or past smoking would truly triple the risk of stroke (RRCD=3.0) then the observed HR in claims data would still be unchanged from the observed 0.79; similarly, if eGFR<45 would truly triple the risk of stroke (RRCD=3.0) then the observed HR in claims data would be 0.83 instead of the observed 0.79 (Figure 2, Panel A). These minimal biases would be counteracted by residual imbalances in INR values although bias estimates are unreliable for INR values due to the high proportion of missing values. The impact on the analysis of major hemorrhage is similar. If HAS-BLED >2 would truly triple the risk of major hemorrhage (RRCD=3.0) then the observed HR in claims data would be 0.77 instead of the observed 0.74 (Figure 2, Panel B).

Figure 2. Quantitative bias analysis based on the observed residual differences in key clinical parameters*, 2a) Stroke: Dabigatran vs. warfarin, assuming HR = 0.79, 2b) Major hemorrhage: Dabigatran vs. warfarin, assuming HR = 0.74,

Figure 2.

Open in a new tab

*Using the observed residual difference in key clinical parameters and assuming a hypothetical observed (apparent) relative risk (aRR) of 0.79 from the claims data analysis these graphs plot the changes in RR for a range of associations between the clinical parameter observed in the EHR data and the hypothetical outcome (RRCD). The RRCD values reach from a non-association of RRCD=1 to strong associations of 3.25 for each clinical parameter. For example, if the duration of AF of 3 year or longer would truly increase the risk of the outcome by 50% (RRCD=1.5) then the observed RR in claims data would truly be 0.82 instead of 0.79 (Panel 2a). Overall the resulting changes are minor under plausible assumptions of RRCD and may even cancel each other out depending on the correlation between the observed clinical parameters, for example the duration of existing AF may be correlated with the chance of being obese. An Excel spread sheet to conduct this analysis is available at: www.drugepi.org/dope-downloads/#Sensitivity Analysis *Stroke 1+ and 5+ years refer to duration since earliest recording of stroke in the EHR prior to treatment initiation *HAS-BLED Score: Labile INR defined as most recent INR <2 or >3 prior to cohort entry

Discussion

From a claims data cohort study of 140,187 NVAF patients newly using DOACs or warfarin that were linked with clinical EHR data in 5,935 (4.2%) patients, we found little evidence of residual confounding by EHR-recorded variables that would meaningfully bias studies on the effectiveness of stroke prevention or incidence of major hemorrhage. Furthermore, the 4% patient sample that successfully linked to EHR data was broadly representative of the much larger claims data study population, although moderate differences in terms of age and some resource use were observed. Indeed, because the EHR information considered in this study is derived from the repositories of physician practices that opted to use the GE Centricity EMR system – a choice which is not expected to be associated with the clinical characteristics of the served patient population – the availability of EHR for patients can essentially be considered “at random”. Among the patient sample linked to EHR data, even before 1:1 PS matching, claims-defined characteristics were well balanced between the respective DOAC and the warfarin groups, which was further improved with PS matching. After PS-matching, almost all EHR-defined patient characteristics were well balanced, although it should be recognized that some EHR variables had a significant amount of missing data. Nevertheless, our findings indicate that it is highly unlikely that residual confounding by EHR-recorded variables would meaningfully bias the claims data analysis in the DOAC monitoring program. While this question has previously been examined in the context of studies for glucose-lowering medications17. – with similar conclusions – this is the first time it has been examined in the context of oral anticoagulants.

We explain this reassuring conclusion by the choices made in study design, analysis planning and the specific clinical setting. We restricted the study population to patients with NVAF. It is well described that restriction can be a powerful tool to remedy confounding.18. 19. Additional restriction to new users and active comparators further reduce confounding.19. 20. In a new user active comparator cohort study, all patients are at a point where their treatment is initiated or escalated by starting a new drug treatment after their prescriber has evaluated their disease state and concluded it is time to change the disease management strategy.7. This design, therefore, makes patients in the comparison groups more similar in measured and unmeasured characteristics. We then used PS matching, which allowed us to adjust for a number of claims-defined covariates. Such comprehensive covariate adjustment will balance a long list of observed confounders and proxy measures of unobserved confounders even if outcome events are infrequent.15. This approach has found wide-spread use in healthcare database analyses because proxy measures can be defined in those data and the threat of residual confounding is high.

Implications for future claims data analyses are two-fold. First, this study confirms that with the right study design and analytic strategy, confounding can be well controlled in database studies.21. 22. 23. We recognize that this is context specific and the preference for an active comparison group may not match the clinical question.19. We further want to stress that our conclusion relates to confounding bias and not other biases caused by misclassification of the outcome or exposure. Second, this study illustrates a scalable approach for checking whether sufficient balance was achieved in important but unmeasured clinical parameters. As EHR databases mature in data quality and completeness they also become increasingly linkable to large claims databases. Having a standing mechanism in place to do this type of linkage in representative patient subsets with all electronic data will expedite the process over medical records abstraction.21. We have previously recommended to check covariate balance in the main study and linked subsets before moving forward with the main outcome analysis.24. 25. Findings of such interim validity checks blinded towards the study results will increase decision-makers’ confidence in the eventual study findings.9. It may also lead to the conclusion that a study should not go forth because it cannot produce valid findings, a decision equally important to avoid polluting the literature with “evidence” that is not fit for purpose.25.

Conclusion:

In database studies of anticoagulation for stroke prevention, a new user active comparator design with 1:1 PS matching on many patient characteristics improved balance on important risk factors not available in claims data but measured in EHR, making confounding bias unlikely. Linking EHR data to a subset of patients in a larger claims database study is a worthwhile and scalable strategy for instilling confidence in findings from database studies.

Methods

Dabigatran monitoring program

This study was conducted in the context of a multi-year program to monitor the safety and effectiveness of dabigatran (NCT02081807, EUPAS5855). The primary outcomes of the monitoring program were stroke and major hemorrhage. The monitoring program involved repeated outcome evaluations over time, starting on October 1, 2010, coinciding with the marketing of dabigatran for stroke prevention in patients with NVAF in the US, and ended in September 2015.

Data source

From the Truven MarketScan healthcare database,26. a longitudinal claims database of commercial U.S. health plans including patients enrolled in Medicare Advantage plans, employer sponsored coverage of seniors, and Medicare supplemental insurance. The MarketScan database contains patient-level information on demographics, health plan enrollment status, records of reimbursed medical services, including inpatient and outpatient encounters with diagnosis and procedure information, and dispensed prescription drugs. For a subset of the population, claims data were linkable to EHRs from select clinics and other outpatient settings providing care to MarketScan beneficiaries. The Institutional Review Board of the Brigham and Women’s Hospital approved the study and signed licensing agreements for use of the Truven data were in place.

Formation of claims-based study population

The study population included three pairwise cohorts of patients aged 18 years or older who initiated dabigatran versus warfarin, rivaroxaban versus warfarin, or apixaban versus warfarin, between October 1, 2010 and December 31, 2014. Patients entered the cohort on the day of a first filled prescription of any of the drugs above defined for each pair-wise cohort as no prior use of any anticoagulant in the previous twelve months and were required to have at least 12 months of continuous enrollment before cohort entry.

We restricted the cohort to patients with a diagnosis of NVAF, defined as an inpatient or outpatient ICD-9 CM diagnosis code of 427.31 at any point prior to drug initiation. We excluded patients with a preexisting diagnosis of valvular comorbidity, CHA2DS2-VASc score <1, or a nursing home admission in the previous twelve months.

Claims-EHR linkage and EHR-based clinical characteristics

For a subset of patients enrolled in the claims data study, insurance claims were enriched with additional data obtained through linkage with EHRs. EHR information was contributed by select clinics and other outpatient settings providing care to MarketScan beneficiaries. Probabilistic linkage was performed by Truven Health Analytics® to preserve patient privacy, including features such as year of birth, sex, three-digit ZIP code, and dates of office visits.27. EHR-defined covariates were captured prior to cohort entry and included, amongst others, health behaviors (smoking status and BMI), duration of AF (the earliest record for an AF diagnosis in the EHR), laboratory test results (INR, estimated glomerular filtration rate [eGFR]),28. and the HAS-BLED score. The HAS-BLED was computed including labile INR defined as the most recent INR <2 or >3 prior to cohort entry. In a sensitivity analysis, we computed the HAS-BLED score in the subgroup of patients with complete information on all HAS-BLED components. If multiple recording of EHR-defined covariates were available, we considered the value closest to the day of cohort entry (see eTable2 for a comprehensive list of the EHR-defined variables). Binary EHR variables capturing the presence or absence of a condition were considered to be truly absent if not recorded in the EHR.

PS-matching within Claims-EHR linked subset

To control for imbalances in patient characteristics between treatment groups in the EHR-linked subset, in three separate multivariable logistic regression models we estimated exposure propensity scores (PS) as the predicted probability of receiving the treatment of interest, i.e. dabigatran, rivaroxaban, or apixaban vs. warfarin, conditional on 78 claims-defined baseline characteristics (Table 1),29. identified during the twelve months before and including the cohort entry date. Emphasis was placed on the identification of claims-defined proxies of stroke and bleed risks, including the HAS-BLED score, the CHA2DS2-VASc score, and prior history of stroke or bleeding. Other patient characteristics included demographics, presence of other comorbidities, use of medications, and indicators of health care utilization as proxy for overall disease state and care intensity. Comorbidities were defined using ICD-9 codes and CPT-4 codes. Exposure groups were 1:1 matched on their PS using nearest neighbor matching without replacement with a maximum caliper of 0.05.30. The PS was re-estimated every 6 months with matching performed within calendar quarters to account for quickly changing prescribing behaviors of newly marketed medications over time.31.

Statistical analysis

To assess whether the claims-EHR linked subset was representative of the overall study population, we compared claims-defined characteristics in the study population for whom EHR data were available and patients without available EHR data using absolute standardized differences (aSD). To assess the potential of residual confounding after PS matching on claims-based variables, caused by clinical characteristics unobserved in claims data, we cross-tabulated the EHR-defined covariates by exposure groups and evaluated imbalances by computing aSD. Meaningful imbalances were defined as aSD greater than 0.1.23.

We quantified the potential bias associated with any residual imbalances in EHR-defined variables based on realistic scenarios of varying exposure-outcome and confounder-outcome associations.32. 33. Findings from these bias models were applied to dabigatran-stroke (HR= 0.79) and dabigatran-major hemorrhage associations (HR=0.74) observed in the monitoring system.12. 34.

Supplementary Material

Supp info

Study Highlights

What is the current knowledge on the topic?

Claims data studies of comparative effectiveness and safety of anticoagulants are often criticized because of the lack of information on critical clinical characteristics, such as underlying bleeding risks, renal function, over-the-counter (OTC) aspirin use, body mass index (BMI), or smoking. Such criticisms could be refuted if the factors unmeasured in claims data studies were in fact balanced between treatment groups when measured in clinical data repositories, due to study design choices and high-dimensional proxy adjustment

What question did this study address?

With the wide-spread use of electronic medical records, subsets of patients identified in administrative claims data can be successfully linked to electronic health records (EHR), and the balance of clinical parameters not documented in claims can be assessed across exposure groups. We sought to evaluate the extent to which balance in unmeasured patient characteristics was achieved in claims data studies, by comparing against detailed clinical information available in EHR data.

What does this study add to our knowledge?

In the context of database studies of anticoagulation for stroke prevention, a new user active comparator design with 1:1 propensity score matching on many patient characteristics improved balance on important clinical risk factors not available in claims data, making confounding bias unlikely

How might this change clinical pharmacology or translational science?

Our manuscript provides evidence that linking EHR data to a subset of patients in a larger claims database study is a worthwhile and scalable strategy for instilling confidence in findings from database studies.

Acknowledgements:

We thank Debra Irwin, Paul Juneau and Kristin Evans for their support and input at various stages of this research.

Funding: This research was supported by a research contract from Boehringer-Ingelheim to the Brigham and Women’s Hospital. The research contract granted Brigham and Women’s Hospital rights to publish all findings as well as final wording of the manuscript.

This research was supported by a research grant from Boehringer Ingelheim.

SS is consultant to WHISCON, LLC and to Aetion, Inc., a software manufacturer of which he also owns equity. He is principal investigator of investigator-initiated grants to the Brigham and Women’s Hospital from Bayer, Genentech and Boehringer Ingelheim unrelated to the topic of this study.

KH reports grant support from the National Institute of Mental Health, and is investigator on grants to the Brigham and Women’s Hospital from Eli Lilly, GlaxoSmithKline and Pfizer unrelated to the topic of this study

Footnotes

CG, JF and JL have no relevant disclosures

DB is an employee at BI X GmbH

KZ and LRF are employed at Boehringer Ingelheim International GmbH

Conflict of Interest:

References

  • 1.Ageno W, Gallus AS, Wittkowsky A, Crowther M, Hylek EM, Palareti G. Oral anticoagulant therapy: Antithrombotic Therapy and Prevention of Thrombosis, 9th ed: American College of Chest Physicians Evidence-Based Clinical Practice Guidelines. Chest. 2012;141(2 Suppl):e44S–e88S. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.You JJ, Singer DE, Howard PA, et al. Antithrombotic therapy for atrial fibrillation: Antithrombotic Therapy and Prevention of Thrombosis, 9th ed: American College of Chest Physicians Evidence-Based Clinical Practice Guidelines. Chest. 2012;141(2 Suppl):e531S–e575S. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Connolly SJ, Ezekowitz MD, Yusuf S, et al. Dabigatran versus warfarin in patients with atrial fibrillation. N Engl J Med. 2009;361(12):1139–51. [DOI] [PubMed] [Google Scholar]
  • 4.Patel MR, Mahaffey KW, Garg J, et al. Rivaroxaban versus warfarin in nonvalvular atrial fibrillation. N Engl J Med. 2011;365(10):883–91. [DOI] [PubMed] [Google Scholar]
  • 5.Granger CB, Alexander JH, McMurray JJ, et al. Apixaban versus Warfarin in Patients with Atrial Fibrillation. N Engl J Med. 2011. [DOI] [PubMed] [Google Scholar]
  • 6.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. [DOI] [PubMed] [Google Scholar]
  • 7.Schneeweiss S A basic study design for expedited safety signal evaluation based on electronic healthcare data. Pharmacoepidemiol Drug Saf. 2010;19(8):858–868. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Schneeweiss S, Seeger JD, Jackson JW, Smith SR. Methods for comparative effectiveness research/patient-centered outcomes research: from efficacy to effectiveness. J Clin Epidemiol. 2013;66(8 Suppl):S1–4. [DOI] [PubMed] [Google Scholar]
  • 9.Schneeweiss S Improving therapeutic effectiveness and safety through big healthcare data. Clin Pharmacol Ther. 2016;99(3):262–265. [DOI] [PubMed] [Google Scholar]
  • 10.Schneeweiss S, Huybrechts K, Gagne J. Interpreting the quality of health care database studies on the comparative effectiveness of oral anticoagulants in routine care. Comp Eff Res. September 2013:33. [Google Scholar]
  • 11.Graham DJ, Reichman ME, Wernecke M, et al. Cardiovascular, bleeding, and mortality risks in elderly Medicare patients treated with dabigatran or warfarin for nonvalvular atrial fibrillation. Circulation. 2015;131(2):157–164. [DOI] [PubMed] [Google Scholar]
  • 12.Seeger JD, Bykov K, Bartels DB, Huybrechts K, Zint K, Schneeweiss S. Safety and Effectiveness of Dabigatran and Warfarin in Routine Care of Patients with Atrial Fibrillation: Thromb Haemost. 2015;114(6):1277–1289. [DOI] [PubMed] [Google Scholar]
  • 13.Graham DJ, Reichman ME, Wernecke M, et al. Stroke, Bleeding, and Mortality Risks in Elderly Medicare Beneficiaries Treated With Dabigatran or Rivaroxaban for Nonvalvular Atrial Fibrillation. JAMA Intern Med. 2016;176(11):1662–1671. [DOI] [PubMed] [Google Scholar]
  • 14.Potpara TS. Dabigatran in “real-world” clinical practice for stroke prevention in patients with non-valvular atrial fibrillation. Thromb Haemost. 2015;114(6):1093–1098. [DOI] [PubMed] [Google Scholar]
  • 15.Schneeweiss S, Rassen JA, Glynn RJ, Avorn J, Mogun H, Brookhart MA. High-dimensional propensity score adjustment in studies of treatment effects using health care claims data. Epidemiology. 2009;20(4):512–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Ray WA. Evaluating medication effects outside of clinical trials: new-user designs. Am J Epidemiol. 2003;158(9):915–920. [DOI] [PubMed] [Google Scholar]
  • 17.Patorno E, Gopalakrishnan C, Franklin JM, et al. Claims-based studies of oral glucose-lowering medications can achieve balance in critical clinical parameters only observed in electronic health records. Diabetes Obes Metab. December 2017. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Rothman KJ, Greenland S, Lash TL. Modern Epidemiology. Lippincott Williams & Wilkins; 2008. [Google Scholar]
  • 19.Schneeweiss S, Patrick AR, Stürmer T, et al. Increasing levels of restriction in pharmacoepidemiologic database studies of elderly and comparison with randomized trial results. Med Care. 2007;45(10 Supl 2):S131–142. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Johnson ES, Bartman BA, Briesacher BA, et al. The incident user design in comparative effectiveness research. Pharmacoepidemiol Drug Saf. 2013;22(1):1–6. [DOI] [PubMed] [Google Scholar]
  • 21.Eng PM, Seeger JD, Loughlin J, Clifford CR, Mentor S, Walker AM. Supplementary data collection with case-cohort analysis to address potential confounding in a cohort study of thromboembolism in oral contraceptive initiators matched on claims-based propensity scores. Pharmacoepidemiol Drug Saf. 2008;17(3):297–305. [DOI] [PubMed] [Google Scholar]
  • 22.Schneeweiss S, Glynn RJ, Tsai EH, Avorn J, Solomon DH. Adjusting for unmeasured confounders in pharmacoepidemiologic claims data using external information: the example of COX2 inhibitors and myocardial infarction. Epidemiol Camb Mass. 2005;16(1):17–24. [DOI] [PubMed] [Google Scholar]
  • 23.Austin PC. Balance diagnostics for comparing the distribution of baseline covariates between treatment groups in propensity-score matched samples. Stat Med. 2009;28(25):3083–3107. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Franklin JM, Rassen JA, Ackermann D, Bartels DB, Schneeweiss S. Metrics for covariate balance in cohort studies of causal effects. Stat Med. 2014;33(10):1685–1699. [DOI] [PubMed] [Google Scholar]
  • 25.Franklin JM, Schneeweiss S. When and How Can Real World Data Analyses Substitute for Randomized Controlled Trials? Clin Pharmacol Ther. 2017;102(6):924–933. [DOI] [PubMed] [Google Scholar]
  • 26.Hansen L The Truven Health MarketScan Databases for life sciences researchers. 2017. [Google Scholar]
  • 27.Huse DM. Linking insurance claims and medical records for outcome research. White paper available at: Dan.Huse@truvenhealth.com
  • 28.Rule AD, Larson TS, Bergstralh EJ, Slezak JM, Jacobsen SJ, Cosio FG. Using serum creatinine to estimate glomerular filtration rate: accuracy in good health and in chronic kidney disease. Ann Intern Med. 2004;141(12):929–937. [DOI] [PubMed] [Google Scholar]
  • 29.Rubin DB. Estimating causal effects from large data sets using propensity scores. Ann Intern Med. 1997;127(8 Pt 2):757–763. [DOI] [PubMed] [Google Scholar]
  • 30.Rassen JA, Shelat AA, Myers J, Glynn RJ, Rothman KJ, Schneeweiss S. One-to-many propensity score matching in cohort studies. Pharmacoepidemiol Drug Saf. 2012;21 Suppl 2:69–80. [DOI] [PubMed] [Google Scholar]
  • 31.Gopalakrishnan C, Huybrechts KF, Shash D, et al. Characteristics of Patients Initiating Oral Anticoagulants in Routine Care: Pharmacoepidemiol Drug Saf. 2015;24:405–406. [Google Scholar]
  • 32.Schneeweiss S Sensitivity analysis and external adjustment for unmeasured confounders in epidemiologic database studies of therapeutics. Pharmacoepidemiol Drug Saf. 2006;15(5):291–303. [DOI] [PubMed] [Google Scholar]
  • 33.Psaty BM, Kuller LH, Bild D, et al. Methods of assessing prevalent cardiovascular disease in the Cardiovascular Health Study. Ann Epidemiol. 1995;5(4):270–277. [DOI] [PubMed] [Google Scholar]
  • 34.Gopalakrishnan C, Huybrechts KF, Ortiz AS, Zint K, Bartels DB, Schneeweiss S. Abstract 16498: Monitoring of Safety and Effectiveness of Dabigatran Relative to Warfarin in Routine Care. Circulation. 2017;136(Suppl 1):A16498. [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supp info
NIHMS993984-supplement-Supp_info.pdf (774.8KB, pdf)

RESOURCES