Efficient estimation of average treatment effects with unmeasured confounding and proxies

One approach to estimating the average treatment effect in binary treatment with unmeasured confounding is the proximal causal inference, which assumes the availability of outcome and treatment confounding proxies. The key identifying result relies on the existence of a so-called bridge function. A parametric specification of the bridge function is usually postulated and estimated using standard techniques. The estimated bridge function is then plugged in to estimate the average treatment effect. This approach may have two efficiency losses. First, the bridge function may not be efficiently estimated since it solves an integral equation. Second, the sequential procedure may fail to account for the correlation between the two steps. This paper proposes to approximate the integral equation with increasing moment restrictions and jointly estimate the bridge function and the average treatment effect. Under sufficient conditions, we show that the proposed estimator is efficient. To assist implementation, we propose a data-driven procedure for selecting the tuning parameter (i.e., number of moment restrictions). Simulation studies reveal that the proposed method performs well in finite samples, and application to the right heart catheterization dataset from the SUPPORT study demonstrates its practical value.