Instrumental variable analysis in the context of dichotomous outcome and exposure with a numerical experiment in pharmacoepidemiology

Abstract

Background: In pharmacoepidemiology, the prescription preference-based instrumental variables (IV) are often used with linear models to solve the endogeneity due to unobserved confounders even when the outcome and the endogenous treatment are dichotomous variables. Using this instrumental variable, we proceed by Monte-Carlo simulations to compare the IV-based generalized method of moment (IV-GMM) and the two-stage residual inclusion (2SRI) method in this context. Methods: We established the formula allowing us to compute the instrument's strength and the confounding level in the context of logistic regression models. We then varied the instrument's strength and the confounding level to cover a large range of scenarios in the simulation study. We also explore two prescription preference-based instruments. Results: We found that the 2SRI is less biased than the other methods and yields satisfactory confidence intervals. The proportion of previous patients of the same physician who were prescribed the treatment of interest displayed a good performance as a proxy of the physician's preference instrument. Conclusions: This work shows that when analysing real data with dichotomous outcome and exposure, appropriate 2SRI estimation could be used in presence of unmeasured confounding.