Vice President, Research OM1, Inc. Tufts University School of Medicine Boston, United States
Background: Understanding the comparative effectiveness and safety of non-vitamin K antagonist oral anticoagulants (NOACs) for patients with atrial fibrillation (AF) in real-world settings is important. Instrumental variable (IV) approaches that leverage provider prescribing preference (PPP) are a promising approach to compare drug effects in the setting of unmeasured confounding. However, the impact PPP definitions have on the validity of the IV is not well understood.
Objectives: To evaluate the strength and validity of an IV using different definitions of provider prescribing preference.
Methods: Data were derived from the OM1 Real World Data Cloud (OM1, Boston, MA), a multi-source real-world data network with linked healthcare claims, social determinants, and electronic medical records data on patients with AF in the United States, and a linked dataset on providers. In this analysis, PPP for index rivaroxaban over apixaban for patients with AF was assessed using the NOAC choice over the last 5, 10 and 20 patients. Logistic regression was used to assess the strength of the instrument with odds ratios (OR) and 95% CIs. IV validity was explored by examining the association between the IV and patient prognostic characteristics within a year of index (date of first observed treatment with apixaban or rivaroxaban) using descriptive statistics.
Results: A total of 182,071 AF patients were identified and linked with 47,758 providers associated with the index NOAC prescription. When the PPP was defined using the previous 5 patients, a total of 95,053 patients and 4,866 providers were included and the instrument was strongly associated with treatment with rivaroxaban versus apixaban (OR=9.7, 95% CI 9.2-10.2). When PPP was defined using the prior 10 patients, fewer providers were available for analysis (61,115 patients, 1,726 providers); the instrument strength improved (OR=17.9, 95% CI 16.6-19.3). Similarly with PPP defined based on the prior 20 patients, the IV was strong (OR=23.0, 95% CI 20.6-25.6) yet included fewer providers and patients (37,283 patients, 666 providers).
The distribution of patient characteristics including age, sex, race, and ethnicity, and BMI was similar between categories of the instrument. Variables strongly associated with key outcomes - Charlson comorbidity index, CHA2DS2-VASc score, and modified HAS-BLED score - were also similarly distributed across levels of the instrument for IVs defined by 5, 10 and 20 patients.
Conclusions: Defining an IV based on provider prescribing preferences is sensitive to the number of patients used to define preference. Use of fewer patients to define preference increased the sample size but was associated with reductions in instrument strength.
