High-dimensional causal inference

Abstract

We present recent progress on estimating bounds on causal effects from observational data, when assuming that these data are generated from an unknown directed acyclic graph. In particular, we present the IDA algorithm for this purpose. IDA is computationally feasible and consistent for high-dimensional sparse systems with many more variables than observations. We validated IDA in biological systems, and will present results on a yeast gene expression data set. Finally, we discuss possible instability issues in high-dimensional settings, as well as extensions towards allowing for hidden variables and predicting the effect of multiple simultaneous interventions.