Imagine you run a massive clinical trial in New York to measure how a new drug affects patient recovery. After careful randomization, you obtain a precise estimate of the drug's causal effect. Now, a hospital in London wants to deploy this drug. Can they simply use your results to predict how their own patients will respond?
In most cases, the answer is no. The London population might have a different age distribution, different baseline health risks, or subtle variations in clinical protocols. Naively merging or transferring these datasets introduces severe bias, as noted by Moore et al. and Bica et al..
Standard machine learning handles this discrepancy via transfer learning or domain adaptation, which primarily focuses on statistical shifts (e.g., how the distribution of patient traits $P(X)$ changes). However, causal inference requires something deeper. We need to know whether the underlying causal mechanisms remain invariant across environments, and if they do not, whether we can mathematically license the transfer of an experimental intervention ($P(y | do(x))$) from a source to a target. This challenge is known as causal transportability.
The Core Intuition: Selection Diagrams
To reason about why and where environments differ, we cannot rely on data alone; we need a formal map of the causal structure. Bareinboim et al. introduced a graphical mechanism called selection diagrams to solve this problem.
A selection diagram is a standard causal Directed Acyclic Graph (DAG) augmented with special indicators called selection nodes (often denoted as $S$). A selection node points directly to any variable whose underlying mechanism varies between the source and target domains.
- If London doctors assign treatments based on a different protocol than New York doctors, an $S$ node points to the treatment variable $X$.
- If the baseline risk of a disease varies due to unobserved genetic differences between the populations, an $S$ node points to the outcome variable $Y$.
[Source Domain] ---> (Invariance Graph) <--- [Target Domain]
|
[Selection Node S]
This framing transforms transportability into a graph-separation problem. The core intuition is elegant: if we can use the rules of causal calculus to mathematically decouple our target query from the selection nodes ($S$), the causal effect is transportable. If the graph structure prevents us from shielding our query from the influence of $S$, then the causal effect cannot be safely transferred without making further parametric assumptions.
The Mechanics: Licensing the Transfer via Do-Calculus
To transport an experimental result from a source domain $\Pi$ to a target domain $\Pi^*$, we want to express the target distribution $P^*(y | do(x))$ using only two things:
- Observational data from the target domain ($P^*(v)$)
- Experimental data from the source domain ($P(v | do(x))$)
In a selection diagram, the target domain is simply the network conditioned on the presence of the selection nodes ($S$), meaning $P^*(y | do(x)) = P(y | do(x) | S)$.
As shown by Bareinboim et al., we can apply Judea Pearl’s do-calculus—a set of three formal rules for manipulating conditional probabilities involving interventions—to transform this expression. The goal is to perform algebraic substitutions that either eliminate $S$ from the experimental expressions or reduce them to purely observational terms where $S$ can be safely observed.
For example, if the graph shows that $S$ is independent of $Y$ given $X$ under an intervention on $X$, do-calculus allows us to drop $S$ entirely from the equation:
$$P(y | do(x), S) = P(y | do(x))$$
This simple equation represents a profound physical reality: it means the experimental effect measured in the source domain is directly identical to the effect in the target domain, despite any other observational differences between the two populations.
Fusing Fragmented Knowledge
In the real world, we rarely enjoy the luxury of a single source domain with perfect experimental data. Instead, our knowledge is often fragmented across multiple environments with limited control.
Limited Control ($z$-Transportability)
What if we cannot directly manipulate the primary treatment $X$ in our source domain due to ethical or logistical constraints, but we can experiment on a subset of proxy or mediator variables $Z$? This scenario is known as $z$-transportability.
Bareinboim et al. and Lee et al. proved that surrogate experiments can still unlock the target causal effect. By mapping how the manipulated mediators $Z$ transmit forces to the outcome $Y$, the authors developed algorithms (such as $sID^z$) that combine target observations with limited source experiments, provided the graph avoids a specific confounding structure known as a "zs-hedge."
Multiple Sources ($m$-Transportability)
What if we have data from multiple different environments, each with its own unique population shifts and experimental limitations? For instance, Domain A experimented on variable $X$, Domain B experimented on variable $Z$, and we want to predict the joint effect in Target Domain C.
Lee et al. and Bareinboim et al. extended the framework to $m$-transportability. Here, graph-theoretic algorithms act as a compiler. They analyze the selection diagrams for all available environments simultaneously, determining if the fragmented experimental pieces can be stitched together like a puzzle to uniquely identify the target query.
Crucially, Bareinboim et al. proved that these algorithmic conditions are complete. If the algorithm fails to find a transport formula for a given graph structure, it is mathematically impossible to identify the causal effect without making stronger parametric assumptions.
Statistical Invariance vs. Causal Power
Why do some metrics transport naturally while others fail? Meder et al. demonstrate this distinction by analyzing how medical trials report risk. They show that Absolute Risk Reduction (ARR) changes drastically depending on the baseline risk of the target population, making it highly non-transportable. Conversely, Relative Risk Reduction (RRR) maps directly onto the structural concept of "causal power." Because RRR isolates the stable, invariant property of the treatment mechanism itself, it transports reliably across environments with wildly different baseline risks.
When structural invariance is imperfect, statistical regularities can bridge the gap. For instance, when combining massive but biased observational data with small, unbiased samples, frameworks developed by Guo et al. and Gu et al. use control variates and statistical shrinkage. These methods adaptively weight data sources, isolating the transportable structural components while filtering out site-specific selection biases.
Advanced Horizons: Soft Interventions and Counterfactuals
The frontier of transportability research moves beyond "atomic" interventions (forcing a variable to take a single fixed value) to handle more complex, real-world queries.
Soft and Stochastic Policies
Real-world policies rarely freeze a variable completely; instead, they alter its probability distribution (e.g., changing a tax rate or updating a medical guideline). Correa et al. and Correa et al. extended transportability theory to these soft interventions, providing a complete framework for mapping how local mechanism shifts propagate through a network under stochastic policies.
Counterfactual Transportability
The deepest layer of causal inference involves counterfactuals—answering retrospective "what-if" questions for specific individuals (e.g., "Would this specific patient have survived if we had administered a different treatment?").
While Lee et al. resolved how to transport conditional causal effects for sub-populations, Correa et al. established the foundations for counterfactual transportability. Their work provides a sound and complete algorithm determining exactly when structural equations from disparate source domains contain enough shared, invariant information to license individual-level counterfactual reasoning in a completely new target environment.
Key Foundational Work
Causal Transportability of Experiments on Controllable Subsets of Variables: z-Transportability (2013) — contributed 4 concept elements to this direction.
Causal Transportability with Limited Experiments (2013) — contributed 3 concept elements to this direction.
m-Transportability: Transportability of a Causal Effect from Multiple Environments (2013) — contributed 2 concept elements to this direction.