Heterogeneous treatment effect estimation with instrumental variable methods

Instrumental variable (IV) methods are widely used to address unmeasured confounding in observational studies but typically estimate the local average treatment effect (LATE) for compliers. If there exist effect modifying variables that also affect the choice of treatment, the LATE may have limited clinical relevance. To address this limitation, we propose a novel framework for estimating the conditional average treatment effect (CATE) for a categorical treatment variable in the presence of both observed confounding and effect-modifying variables in addition to valid IVs. Building on recent developments in data fusion between randomized trial and observational data, we combine IV estimators with estimators based on covariate adjustment. By making parametric assumptions about the residual unobserved confounding effect, we obtain a consistent estimator of the CATE with relatively high efficiency. In a simulation study with nonlinear effect modification and unobserved confounding, we show that our estimator outperforms competing IV and covariate-adjusted estimators in terms of mean squared error. We further illustrate the applicability of the method, utilizing it to compare tumor necrosis factor alpha inhibitors as the first line biologic treatment of rheumatoid arthritis conditional on age. The analysis illustrates how the proposed method can estimate population-relevant, clinically interpretable treatment effects and suggests smaller differences between treatments among older patients. This work provides a general framework for combining instrumental and confounding information to estimate individualized causal effects when treatment effects are heterogeneous and there are more than two treatment alternatives.