Background: Negative control outcome (NCO) models are increasingly used to detect unmeasured confounding in observational studies. It is unclear how to incorporate multiple NCO results into a summary assessment of possible residual bias.
Objectives: To assess unmeasured confounding in a comparison of migraine preventive medications by using a Bayesian approach to estimate the distribution of possible bias terms derived from multiple NCO model results.
Methods: US-based MarketScan® Commercial and Medicare Databases were used to assess the comparability of patients initiating erenumab to patients initiating other calcitonin gene-related peptide (CGRP) pathway antagonists between May 2018 and June 2021. Inverse-probability treatment weighting was used to balance treatment groups with respect to baseline characteristics. Cumulative risk differences (RDs) for 11 NCOs at various time points between 0.5 and 2.5 years of follow-up were calculated for intention-to-treat (ITT) and on-treatment (OT) models. A Bayesian model using the bootstrapped effect estimates for these 11 NCOs as inputs was used to generate a bias parameter.
Results: After weighting, 19,220 erenumab and 23,244 non-erenumab CGRP pathway antagonist new users had comparable baseline characteristics (each standardized mean difference < 0.10). RDs (95% confidence intervals) for NCOs over 2.5 years of ITT follow-up comparing erenumab versus non-erenumab CGRP pathway antagonists were -1.7% (-3.2%, -0.2%) for osteoarthritis diagnoses, -1.4% (-2.7%, -0.1%) for anemia diagnoses, -0.3% (-1.3%, 0.8%) for accident diagnoses, -0.3% (-1.1%, 0.5%) for asthma diagnoses, 0.1% (-0.7%, 0.8%) for fracture diagnoses, -4.4% (-6.2%, -2.6%) for herpes zoster vaccination, -3.5% (-5.1%, -1.8%) for electrocardiogram utilization, -1.8% (-3.3%, -0.2%) for echocardiogram utilization, -1.3% (-3.2%, 0.7%) for pelvic exam utilization in women, 1.1% (-0.4%, 2.6%) for mammography utilization in women, and 2.7% (1.4%, 3.9%) for influenza vaccination. Bayesian bias parameters indicated that for 2.5 years of follow-up, absolute values of bias in RDs had a 90% probability of being less than 0.0460 for ITT models and 0.0639 for OT models. Absolute values of bias in RDs for 0.5 years of follow-up were estimated at a 90% probability to be less than 0.0140 and 0.0152 for ITT and OT models, respectively.
Conclusions: This study demonstrated a Bayesian approach for estimating possible residual bias in comparative studies using multiple NCOs. In this example, the Bayesian analysis suggested unmeasured confounding increased with longer length of follow-up and in OT versus ITT models.
