B1254
Title: Recent advances in causal discovery for time series and optimal adjustment for causal effect estimation
Authors: Jakob Runge - German Aerospace Center (Germany) [presenting]
Abstract: Two methods for statistically optimal causal discovery and causal effect estimation are presented. For the former, in the sense that the conditioning sets in the iterative tests are constructed such as to achieve high effect size and hence high recall. The method is designed for linear and nonlinear, lagged and contemporaneous causal discovery from observational time series in the causally sufficient case with an extension to the case with hidden confounding. For optimal causal effect estimation, a method for selecting optimal backdoor adjustment sets is presented to estimate causal effects in graphical models with hidden and conditioned variables. Previous work has defined optimality as achieving the smallest asymptotic estimation variance and derived an optimal set for the case without hidden variables. For the case with hidden variables, there can be settings where no optimal set exists. A necessary and sufficient graphical criterion is defined for the existence of an optimal adjustment set and a definition and algorithm to construct it. The results translate to minimal asymptotic estimation variance for a class of estimators whose asymptotic variance follows a certain information-theoretic relation. Code is available as part of the Python package Tigramite.
