TL;DR: We develop AMRIV, an adaptive, multiply robust estimator for average treatment effect (ATE) estimation when treatment can only be encouraged via a binary instrument. Our method learns an optimal, covariate-dependent assignment policy that minimizes estimator variance, and pairs it with a sequential influence-function–based estimator that is both semiparametrically efficient and robust to nuisance misspecification. We also provide time-uniform asymptotic confidence sequences and validate the approach in synthetic and semi-synthetic experiments.
Use the following commands to replicate the figures from the "Efficient Adaptive Experimentation with Non-Compliance" paper:
-
For Figure 1 & 2:
python run_synthetic.py -
For Figure 3:
python run_semi_synthetic.py