Title: Detection of instrumental variables in causal models: An algorithmic framework
Authors: Maciej Liskiewicz - University of Luebeck (Germany) [presenting]
Benito van der Zander - University of Luebeck (Germany)
Abstract: Instrumental variables (IVs) are widely used tool to identify causal effects from observational data. For this purpose IVs have to be exogenous, i.e., causally unrelated to all variables in the model except the explanatory variable. Since these IVs do not exist in many model instances, the approach has been generalized to conditional IVs that only require exogeneity conditioned on a set of covariates. Another generalization is to use instrumental sets that allow us to identify causal effects if no single instrument exists. This has led to a wider choice of potential IVs. However, a significant barrier to the applications of this method is of algorithmic nature: So far, it was not clear whether such generalized IVs can be found efficiently. We address two issues with generalized IVs. First, we discuss new natural concepts of IVs, which interpolate between the existing notions. Second, we provide effective algorithms for detection of such IVs and show NP-hardness for the most generalized levels. Together this implies a complete and constructive solution to causal effect identification using IVs in linear causal models.