In an ideal world, causal discovery would begin with a single, massive dataset containing every relevant variable, perfectly randomized and free of bias. In reality, researchers work with fragments: a large but biased observational study here, a small clinical trial there, and perhaps a legacy dataset from a different population that only measured a subset of the variables of interest.
Causal data fusion is the formal framework for "stitching" these disparate pieces together. It allows researchers to answer causal questions that no single dataset could address on its own by leveraging the structural invariants shared across different data-generating environments.
Stitching Fragmented Networks
The most fundamental challenge in data fusion is the problem of overlapping variables. Often, no single study observes all variables $\{A, B, C, D\}$ simultaneously. Instead, one study might observe $\{A, B, C\}$ and another $\{B, C, D\}$.
Tillman et al. demonstrated that it is possible to reconstruct a global causal map from these local fragments using the ION (Integration of Overlapping Networks) algorithm. The intuition is that the causal structure must satisfy the logical independence constraints found in all datasets simultaneously. If Study 1 shows that $A$ and $C$ are independent given $B$, any global graph must respect that, even if Study 2 (which doesn't see $A$) suggests a different relationship between $B$ and $C$.
This logic extends to resolving contradictions. Claassen and Heskes showed that causal discovery can be framed as a process of elimination. By using negative constraints from one experiment—such as ruling out a causal link—researchers can resolve ambiguities in another experiment where the direction of causality was previously unclear.
The Synergy of Observations and Experiments
A recurring theme in this research is the fusion of "cheap but messy" observational data with "expensive but clean" experimental data. Observational data provides scale and captures real-world behavior but is plagued by unobserved confounders. Experimental data (like Randomized Controlled Trials) provides gold-standard causal evidence but is often too small to capture individual-level nuances.
Li and Pearl and Li demonstrated that combining these sources significantly narrows the mathematical bounds on counterfactual probabilities. For example, in the "unit selection problem"—identifying which specific individuals will respond to a treatment—observational data reveals how people naturally behave, while experimental data reveals how they respond to being forced. Fusing them allows practitioners to identify "desirables" (those who respond only when treated) more accurately than A/B testing alone.
Similarly, Gan et al. showed that even a small "deconfounded" subset can act as a lighthouse for a massive confounded dataset. The large dataset constrains the marginal distributions, while the small subset provides the necessary causal anchors to correct for bias, leading to treatment effect estimates that are far more precise than those derived from either source in isolation.
Transportability and Multi-Environment Synthesis
Even when data is clean, it may come from the wrong population. A clinical trial in Los Angeles might not directly apply to a population in New York due to differences in demographics or environmental factors. This is the "transportability" problem.
Bareinboim and Pearl developed the sID algorithm, which provides a formal language for this transfer. By mapping the structural differences between environments (represented as "selection nodes" in a causal graph), the algorithm determines if a causal effect in a target domain can be calculated using a "transport formula" that re-weights the source experimental results.
This concept matures into m-transportability, where researchers pool data from multiple source environments. Lee and Honavar and Bareinboim et al. proved that effects unidentifiable from any single source can become identifiable when sources are combined. One environment might provide the experimental link between $X$ and $Z$, while another provides the link between $Z$ and $Y$; causal data fusion provides the calculus to chain these findings into a single estimate for the effect of $X$ on $Y$.
Correcting for Selection and Confounding Biases
Data fusion also offers a path to recovery when data is systematically corrupted. Tian et al. and Correa and Bareinboim addressed scenarios where data suffers from both hidden common causes (confounding) and non-random sampling (selection bias).
The intuition here is to use unbiased external data—such as census records or a small pilot study—to "re-calibrate" the biased primary dataset. By integrating these sources, the authors show that one can recover the true causal distribution even when the primary data source is fundamentally unrepresentative. This is achieved by identifying which parts of the causal mechanism are invariant across the biased and unbiased samples, allowing the researcher to "adjust out" the selection mechanism.
Unifying Discovery and Estimation
Finally, recent work has moved toward treating all data types—observational, interventional, and counterfactual—as facets of a single estimation task. Magliacane et al. formulated causal discovery as a constraint satisfaction problem where interventions are simply additional logical constraints.
This unified view is particularly powerful in complex domains like reinforcement learning. Forney et al. and Gasse showed that agents can learn faster by fusing their own "trial and error" data with passive observations of other agents. By treating the "intent" of other agents as a proxy for unobserved confounders, these systems turn the "messiness" of observational data into a strategic advantage for causal reasoning.
Key Foundational Work
Integrating Locally Learned Causal Structures with Overlapping Variables (2008) — contributed 4 concept elements to this direction.
Causal discovery in multiple models from different experiments (2010) — contributed 3 concept elements to this direction.