Abstract
Instrumental variable (IV) regression is a standard strategy for learningcausal relationships between confounded treatment and outcome variables byutilizing an instrumental variable, which is conditionally independent of theoutcome given the treatment. In classical IV regression, learning proceeds intwo stages: stage 1 performs linear regression from the instrument to thetreatment; and stage 2 performs linear regression from the treatment to theoutcome, conditioned on the instrument. We propose a novel method, {\it deepfeature instrumental variable regression (DFIV)}, to address the case whererelations between instruments, treatments, and outcomes may be nonlinear. Inthis case, deep neural nets are trained to define informative nonlinearfeatures on the instruments and treatments. We propose an alternating trainingregime for these features to ensure good end-to-end performance when composingstages 1 and 2, thus obtaining highly flexible feature maps in acomputationally efficient manner. DFIV outperforms recent state-of-the-artmethods on challenging IV benchmarks, including settings involving highdimensional image data. DFIV also exhibits competitive performance inoff-policy policy evaluation for reinforcement learning, which can beunderstood as an IV regression task.