Problem
A business has to choose actions by their effect on a target metric, but an A/B test is sometimes very long, very expensive or impossible for physical or ethical reasons. Observational data then has to answer the question. Correlations there mislead when a hidden common cause drives two variables, and the right controls depend on a causal graph that most teams never build. Estimation methods get most of the attention; the graph stage gets the least.
Example from the article: the number of Nobel laureates per country correlates strongly with per-capita chocolate consumption. Neither plausibly causes the other; a third variable such as GDP most likely drives both.
Solution in brief
Split causal inference into two stages: identification (find a causal graph and a correct adjustment set) and estimation (uplift-style models on that set). For identification, use constraint-based discovery. FCI-family algorithms drop the assumption of no hidden confounders and flag suspicious edges. Speed is the practical barrier, and the speaker's team addresses it with the CAIMAN library of parallelized algorithms.
Questions this case answers
- How to estimate the effect of an action from observational data when an experiment is too slow, expensive or impossible, by first finding a causal graph and a valid adjustment set
- How to speed up causal graph discovery so many candidate graphs can be tested when hidden common causes may exist
- How to estimate the effect of actions when A/B tests are too slow, too expensive or impossible
- How to choose the right set of controlled variables before estimating a treatment effect
Tools and methods
- PC algorithm — classic constraint-based causal discovery, assumes no hidden confounders
- FCI, FCI-stable, RFCI, FCI+ — discovery that tolerates hidden confounders and selection bias
- RLSD, SALAD — recent methods that recover the structure of hidden variables
- CAIMAN — library of CPU-optimized discovery algorithms with bootstrapping (B-RAI, B-FCI)
- DoWhy, EconML — effect estimation after identification
Google or a one-time email code. No password.