A0603
Title: Causal inference with hidden confounders: a comparison between two stage least squares and the causal Dantzig
Authors: James Long - University of Texas MD Anderson Cancer Center (United States) [presenting]
Min Jin Ha - UT MD Anderson Cancer Center (United States)
Abstract: The Causal Dantzig (CD) estimates causal effects by exploiting shifts in the exposure distribution across a set of data collection environments (e.g. experimental and observational). It is one of a small number of methods, along with instrumental variables techniques, which are consistent under hidden confounding. We propose a model for jointly analyzing the performance of the CD and the classical Two-Stage Least-Squares (TSLS) instrumental variable estimator. The model is appropriate for many settings, including genetic perturbation experiments and most standard applications of instrumental variables estimators. We derive the first analytic results comparing the CD and TSLS, including conditions under which TSLS has lower asymptotic variance than the CD. We compare regularized versions of the CD, TSLS, and other state-of-the-art methods in high dimensional genetic perturbation simulations and a real yeast perturbation data set. From a fitting perspective, the high dimensional CD is simpler than TSLS because the environment simplifies the selection of tuning parameters. Performance of the methods is assessed by accuracy in predicting the effect of test set knock out experiments. Regularized TSLS procedures obtain the best performance in many scenarios tested.
