Methods
Data
We used the IMS Health LifeLink Health Plan Claims Database. This database contains commercial health plan information from managed care plans and other sources (such as Medicare and Medicaid) throughout the United States, and it is generally representative of the national, commercially insured population in terms of age and sex. The database contains claims files and enrollment files. We used the claims files to derive information on inpatient and outpatient diagnoses documented as the ICD-9-CM (international classification of diseases, ninth revision, clinical modification) codes, procedures as the Current Procedural Terminology (CPT-4) codes or the Healthcare Procedural Coding System (HCPCS), prescriptions as National Drug Code (NDC), date of services, and days of the prescription supplied. We used the enrollment files to derive information about patients' demographic characteristics (including year of birth and sex) and monthly medical/pharmacy enrollment indicators.
Study Population
Our study population included 244 872 enrollees with a prescription of warfarin, dabigatran, or rivaroxaban between 1 October 2010 and 31 March 2012. We entered each person into the cohort at the date of his or her first prescription for any of the three study drugs after 1 October 2010. To be eligible for this study, we required people to be aged 18 years or older, have continuous medical and pharmacy enrollment in the six months before the entry date (the baseline period), have none of the three drugs prescribed in the baseline period (new user design), have the first prescription before 31 March 2012, have known age and sex, and not have a previous bleeding event. The final study sample of 46 163 patients included 4907 dabigatran users, 1649 rivaroxaban users, and 39 607 warfarin users (fig 1).
(Enlarge Image)
Figure 1.
Flow chart of sample size
We defined a patient's observation ending date as the earliest of five dates. The first of these was the last date of (continuous) exposure to the same drug. The last exposure date was the last day of continuous drug at hand plus 14 days; we used these 14 days to take into account the clearance of the drug. For example, if there was a gap of 14 days or less between two consecutive prescriptions of the same drug, we considered the exposure to be continuous. However, if a patient switched to another drug within the 14 day period of the previous exposure, we considered the day before the start of the second drug to be the last exposure date. The second date was the date of the loss of medical or pharmacy enrollment. The third was the end date of the study, which was 31 March 2012. The fourth was the date before the first date of hospital admission not related to gastrointestinal bleeding, because we did not have prescription information during hospital admission. The last date was the first date of gastrointestinal bleeding. If gastrointestinal bleeding occurred during hospital admission, we used the first date of admission as the bleeding date for patients admitted for gastrointestinal bleeding. We censored a patient if the observation ending date was not equal to the first bleeding date.
Outcome and Independent Variables
Our independent variable of interest was the type of drug exposure: warfarin, dabigatran, or rivaroxaban. We used control variables including demographics (age groups, sex, and region), three binary clinical conditions (having any diagnosis of renal failure, trauma, or Helicobacter pylori infection), three binary drug indicators (having any prescription of non-steroidal anti-inflammatory drugs, proton pump inhibitor, or steroid), and six levels of counts of Clinical Classification Software (CCS). The CCS is a tool, developed by the Agency for Healthcare Research and Quality, for clustering patients' diagnoses and procedures into a manageable number of clinically meaningful categories; the higher the CCS score, the higher the comorbidity. All variables were derived from claims data in the baseline period. We incorporated propensity scores into the models as a means of weighting the observations to account for potential confounding by the variables above. Our outcome of interest was gastrointestinal bleeding, which we identified using ICD-9 codes and CPT codes validated in a recent study (web appendix 1).
Propensity Score
To control for the differences in patients' characteristics across warfarin, dabigatran, and rivaroxaban users, we developed two propensity scores: one to predict whether a person used dabigatran relative to warfarin among dabigatran and warfarin users (n=44 514) and another to predict use of rivaroxaban relative to warfarin among rivaroxaban and warfarin users (n=41 256). To develop these propensity scores, we included the aforementioned control variables; the only exception was that we entered CCS into the propensity score model as 285 mutually exclusive binary indicators.
Various methods have been proposed to apply propensity scores, including matching, inverse probability of treatment weighting, stratification, and regression covariates; each has its own advantages and disadvantages, and no single method consistently outperforms other approaches. We used propensity score weighting because we did not want to lose the observations of treated patients (compared with matching) and we wanted one interpretable overall treatment effect (compared with stratification). We applied the average treatment effect of the treated weighting because this allows us to estimate the average effect of treatment on patients who received the treatment; that is, we compared the hazards of gastrointestinal bleeding among dabigatran or rivaroxaban users with the hypothesized situation had they taken warfarin instead of the oral anticoagulant. This is particularly useful when the study sample is likely to differ systematically from the overall population. We compared the balance in baseline covariates before and after weighting by using the standardized difference; we considered a standardized difference less than 0.1 to be a negligible difference between treatment groups. After we applied average treatment effect of the treated weighting, standardized differences of all available covariates between dabigatran and warfarin users and between rivaroxaban and warfarin users were reduced to 0.05 or smaller, suggesting that the groups were well balanced (Appendix 2).
Statistical Analysis
We used χ tests and analysis of variance/Kruskal-Wallis tests to examine whether patients' characteristics were different across the three groups of drug users. We present Kaplan-Meier survival curves of having gastrointestinal bleeding stratified by the three different drug groups (fig 2). We created two separate Cox proportional hazard models with propensity score average treatment effect of the treated weighting to examine the association between anticoagulant exposure ("dabigatran v warfarin" and "rivaroxaban v warfarin") and gastrointestinal bleeding, and we calculated robust estimates of standard errors for all variables in the model. Control variables could enter the model as either regression covariates or stratification factors; if a variable violated the proportional hazard assumption, this variable would enter the model as a stratification factor. We started with all control variables as regression covariates and examined whether the proportional hazard assumption was violated for any control variable at the P=0.1 level (Kolmogorov-type supremum test). Across two models, three variables reached that level—age groups, CCS categories, and having any use of non-steroidal anti-inflammatory drug in the baseline period; these were later included in the model as stratification factors. We re-checked the proportional hazard assumption and found that it was not violated with all remaining variables (P≥0.1). We also did a post hoc analysis stratifying our results for patients above and below the age of 65 years, using the same model as for the total sample. We used SAS version 9.2 for all analyses.
(Enlarge Image)
Figure 2.
Survival function and number of patients at risk over time by drug use
Sensitivity Analyses
We used several sensitivity analyses to examine whether our findings were robust. Firstly, we evaluated two additional models: one including all variables as regression covariates and another including all variables as stratification factors. Secondly, we varied the length of washout period from seven to 30 to 45 days to check the robustness of our results. Thirdly, we censored all inpatient records owing to the lack of the prescription information during hospital admission, as we wanted to examine whether such exclusion would affect our findings. Finally, we additionally included the HAS-BLED (hypertension, abnormal renal/liver function, stroke, bleeding history or predisposition, labile international normalized ratio, elderly, drugs/alcohol concomitantly) bleeding risk score in the model to control for a patient's risk of bleeding and examine whether our results would change. Given that we did not have laboratory data, we excluded the labile international normalized ratio from construction of this risk score.