In causal inference, we often find ourselves in a data-rich but information-poor environment. We might have massive amounts of observational data (the "messy" reality) and a handful of controlled experiments (the "clean" but expensive interventions). Traditionally, researchers treated these as separate silos: you either identify an effect from observations or you run an experiment.
The research direction led by Elias Bareinboim, Juan D. Correa, and Sanghack Lee unifies these silos. It asks a fundamental question: if we have an arbitrary collection of observational and experimental data, can we uniquely determine a specific causal effect? This is the problem of Generalized Identifiability (gID).
The Algebra of Causality
To understand why this works, we must look at the "grammar" of causal reasoning: Pearl’s do-calculus. As established in The mathematics of causal inference by Judea Pearl, do-calculus consists of three rules that allow us to transform probability expressions involving interventions (the "do" operator) into expressions involving only observations.
The breakthrough in this field is the proof of completeness. In Identification of Conditional Interventional Distributions, Ilya Shpitser and Judea Pearl demonstrated that these three rules are sufficient. If a causal effect can be identified from observational data, a sequence of these three rules will eventually find the formula. If the rules cannot find it, the effect is mathematically unidentifiable.
Data Fusion and Transportability
Modern research extends this logic to "Data Fusion." In Identifiability from a Combination of Observations and Experiments, the authors show that do-calculus isn't just for moving from experiments to observations; it is a complete logic for combining any set of experiments.
Imagine you have an experiment on how a drug affects blood pressure in New York, and observational data on how blood pressure affects heart disease in London. Can you infer the effect of the drug on heart disease in Tokyo? This is the problem of General Transportability.
As explored in General Transportability - Synthesizing Observations and Experiments from Heterogeneous Domains, do-calculus remains the complete tool for this task. It allows researchers to "transport" causal knowledge across different populations by formally accounting for the differences in their underlying distributions.
The Limits of Knowledge: Thickets and Positivity
If do-calculus is complete, how do we know when we've failed? Sanghack Lee and colleagues introduced the concept of "thickets" in General Identifiability with Arbitrary Surrogate Experiments. A thicket is a specific graphical structure where causal paths are so entangled with hidden variables that no combination of the available experiments can untangle them.
Furthermore, recent work on conditional generalized identifiability (c-gID) by Yaroslav Kivva, Jalal Etesami, and Negar Kiyavash has refined this by addressing the "positivity assumption." They proved that do-calculus is sufficient even for conditional effects, provided we account for the possibility of dividing by zero in our statistical formulas—a critical detail for practical algorithm implementation.
The Counterfactual Frontier
The most recent evolution of this direction moves from interventions ("What happens if I take this action?") to counterfactuals ("What would have happened if I had acted differently?").
In Counterfactual Graphical Models: Constraints and Inference, Juan D. Correa and Elias Bareinboim introduce Ancestral Multi-world Networks (AMWN) and a new Counterfactual Calculus (ctf-calculus). This framework generalizes do-calculus to reason across multiple parallel worlds simultaneously.
By formalizing principles like consistency (if you do X, X happens) and exclusion (some variables don't affect others), they provide a complete set of algebraic rules for counterfactuals. This means we now have a rigorous, algorithmic way to determine if "what if" questions can be answered from the data we have on hand.