Creating a causal graph from observed data without prior knowledge on causal relationships of the features

Question

Is it possible to create a causal graph from data collected in a survey? If I find lets say feature A and feature B are dependent on each other, is it possible to infer whether A causes B (will result in an edge from A to B in the graph) or B causes A (will result in an edge from B to A in the graph)?

I am currently reading some books on causal inference. Some more help would be highly appreciated

Answer 1

You are looking for "Causal Discovery"

The keyword you are probably looking for is "Causal Discovery" (in your case from observational data). There is too much to say to summarize it in just one post, but here are a few places to get started.

Word of Caution: The one thing you should know about causal discovery, especially from observational data, is that despite the existence of many algorithms and intricate proofs, it tends to rest on extremely strong assumptions that are most likely violated at least to some degree in practice. Despite having done a fair amount of research in the field, I could not give you a truly compelling example of a practical applicaton (I'd be very happy to be shown otherwise).

These are some fairly comprehensive review papers:

• Review of Causal Discovery Methods Based on Graphical Models
• Causal Structure Learning
• A survey of Bayesian Network structure learning

A nice video introduction is given in chapter 10 of the Causal Inference Course Lectures.

The Elements of Causal Inference is a good textbook to learn more.