Treatment effect estimates varied depending on the definition of the provider prescribing preference-based instrumental variables

Abstract

Objective: The instrumental variable (IV) method can remove bias because of unobserved confounding, but it is unclear to what extent the choice of the IV may affect the results. We compared the estimates obtained with different provider-based IVs in a real-life observational comparative drug effectiveness study.

Study design and setting: We assessed the effectiveness of rhythm vs. rate control treatment in reducing 5-years mortality in a population-based cohort of patients with atrial fibrillation. We compared the IV treatment effect estimates obtained from two-stage least square regression models using nine alternative provider-based IVs defined at either hospital or physician level.

Results: All nine IVs reduced the covariate imbalance between the treatment groups. Yet, there were large variations in both the point estimates and the width of the confidence intervals obtained with alternative IVs. Relative to the physician-based IVs, the hospital-based IVs were stronger, had smaller variance, and produced less extreme point estimates.

Conclusions: The IV estimates of treatment effect may vary considerably depending on the IV definition. Choosing the strongest IV could reduce the variance of the IV estimates.