Causal discovery under hidden confounding
Not all causal discovery methods are helpless in the face of hidden confounding.
In this section, we’ll learn about the FCI algorithm, which can operate when some or all confounding variables are unobserved. We’ll implement the FCI algorithm using the causal-learn package and – finally – discuss two more approaches that can be helpful when our dataset contains potential unobserved confounders.
The FCI algorithm
FCI (Sprites et al., 2000; Sprites et al., 2013) is a constraint-based algorithm. This means that the algorithm uses a set of conditional independence tests in order to decide which edges exist and what their orientations are. The FCI algorithm can be thought of as an extension of the PC algorithm that can work on an extended class of graphs, called inducing path graphs. The theory behind inducing path graphs is beyond the scope of our book. If you want to learn more, check out Chapter 6 of Sprites...
