Trevis Litherland’s Post

View profile for Trevis Litherland

Manager, Data Science at LexisNexis

Causality: Chapter 2 – Post 2 ...But what if there are hidden (latent) variables? Pearl extends the IC algorithm to the latent structure case (the IC* Algorithm). The resulting “marked diagrams” outputs have four different types of edges, (described in further detail in the next section): 1)       “Marked,” directed edges (genuine causality) 2)       Unmarked, directed edges (potential causality) 3)       Bi-directed edges (spurious association) 4)       Undirected edges (unknown relationship) The notion of genuine causation is a little simpler once he introduces temporal information, and Pearl explores temporality further via the intriguing topic of “statistical time.” The chapter ends with a defense of the approach described above, both in practical and philosophical terms. Whereas minimality is relatively uncontroversial, DAGs’ inherently Markovian structure and Pearl’s notion of stability have both been challenged. Pearl finds these challenges to be answerable, and so it seems to me. I very much enjoyed this chapter. After all, it’s a very natural question to ask when you begin playing with diagrams: “What sort of DAGs should I be writing down, based on this probability distribution?” While the IC* Algorithm doesn’t always provide a unique answer, it does provide a very clear picture of what causal relationships we can know “for certain” (genuine causation), which ones we can feel good about (potential causation), which ones involve a latent variable (spurious association), and which ones exhibit a dependence that we can’t elucidate further. Overall, this is quite an achievement, one I’m still pondering, as I move on to Chapter Three.

To view or add a comment, sign in

Explore topics