Probabilistic Causal Fusion Methods

• Probabilistic causal fusion is a framework that rigorously integrates multi-source data, structural constraints, and uncertainty quantification for improved causal inference.
• It leverages techniques such as Neyman-orthogonal risk minimization and kernel ridge fusion to achieve robust estimation despite selection bias and confounding.
• Its methodologies incorporate uncertainty quantification and scalable computation, with applications spanning biostatistics, econometrics, epidemiology, and spatiotemporal forecasting.

Probabilistic causal fusion refers to the rigorous integration of data, constraints, and uncertainty quantification across heterogeneous sources—ranging from randomized trials, registries, or opportunistically collected datasets—for improved causal identification, estimation, and inference. Core to its methodology is the fusion of both probabilistic (distributional) information and causal (structural, interventional) information, with formal guarantees on efficiency, robustness, and validity under non-ideal conditions such as limited source alignment or differential selection. Recent research has produced identification algorithms, minimax-optimal estimators, uncertainty quantification frameworks, and scalable computational tools for probabilistic causal fusion across domains including biostatistics, econometrics, epidemiology, and spatiotemporal forecasting.

1. Foundational Setup: Fusing Data, Structure, and Uncertainty

Probabilistic causal fusion formalizes the integration of multi-source data as an inference task in a superpopulation described by potential outcomes or a structural causal model (SCM). For concreteness, consider the estimation of a target causal parameter such as a dose-response function $\theta_0(a) = \mathbb{E}[Y^a]$ in a population $Q^0$, where observations are only available from multiple sources partially aligned with $Q^0$ due to disparities in covariate or outcome distributions. Each unit is labelled with a source indicator $S\in\{1,\ldots,k\}$, indicating (possibly partially) aligned distributions for covariates $X$ or conditional distributions for $Y|(X,A)$ (e.g., "X-aligned" or "Y-aligned" sources) (Lim et al., 21 Oct 2025).

The identification challenge is to express $\theta_0(a)$, or more general interventional distributions, as a functional of observed data-generating processes $P^{(s)}$ from possibly distinct source populations, along with the structural knowledge encoded in a causal DAG or potential outcome model. Graphical models, including labelled-selection DAGs and acyclic directed mixed graphs (ADMGs), are often used to track selection mechanisms, context-specific constraints, and to precisely define which statistical and counterfactual queries are estimable (Lee et al., 9 Apr 2024).

2. General Identification and Partial Identification under Systematic Selection

Rigorous probabilistic causal fusion must address nontrivial selection, missing data, and confounding structures. Under the LS-ADMMG formalism (Lee et al., 9 Apr 2024), selection into datasets is modelled via an explicit selector variable $S$ whose values index the available regimes or interventions. The SS-ID algorithm establishes a general approach: any target $p(Y(a),S=\emptyset)$ can be expressed as a sum over districts (collections of variables sharing unmeasured confounders) of kernels $p(D^*|do(pa_G(D^*)))$, using fixing operators, recursive graphical criteria, and context-specific independence properties. In the presence of systematic selection (beyond SCAR/SAR), partial completeness is established: if the algorithm fails at positivity, hedge, or thicket conditions, the effect is nonidentified.

The method is sound for arbitrary unobserved confounding and selection, and reduces to standard ID and gID when selection is untangled or only involves observed parents. The formalism elegantly adapts Rubin's hierarchy (MCAR/MAR/MNAR) into a causal-selection taxonomy (SCAR/SAR/SNAR), generalizing identification beyond simple missing data (Lee et al., 9 Apr 2024).

3. Neyman-Orthogonal Risk and Kernel Ridge Fusion for Dose-Response

Estimating functional causal parameters—such as continuous dose-response curves—under data fusion is facilitated by formulating an orthogonal empirical risk minimization. A Neyman-orthogonal loss function, constructed via influence-function analysis, ensures that estimation errors in nuisance parameters enter at higher order, conferring insensitivity to imperfect nuisance learning (Lim et al., 21 Oct 2025). Specifically, for $\theta_0(a)$ in a fusion regime, the risk

$$L_{Q^0}(\theta) = \int (\theta(a) - \theta_0(a))^2 \mu(da)$$

is estimated by a one-step (bias-corrected) loss that combines information from both X- and Y-aligned sources via appropriate weighting and kernel regression. Stochastic approximation replaces intractable integrals by Monte Carlo sampling, preserving Neyman-orthogonality up to $o_p(1)$, and yielding a closed-form kernel ridge regression solution in a chosen RKHS.

Explicitly, the estimator is computed as follows:

• Compute source-specific pseudo-outcomes $u_i$ and $v_i$ using estimated nuisance functions (marginals, regressions, density ratios).
• Assemble a block Gram matrix over exposures (true and sampled).
• Solve for coefficients $(\hat\beta, \hat\gamma)$ in a regularized linear system, yielding

$$\hat\theta(\cdot) = \frac{1}{\sqrt{n}} \sum_{j=1}^n [\hat\beta_j K(\cdot,a_j) + \hat\gamma_j K(\cdot, b_j)].$$

Oracle-type upper bounds show that fusion always yields at least as fast a rate as the best-aligned single source, with strict minimax efficiency gains in nontrivial partial-overlap regimes (Lim et al., 21 Oct 2025).

4. Robust Estimation, Double Robustness, and Efficiency

Direct learning approaches for multi-source fusion formulate pseudo-outcomes and weighted regression such that the target CATE function $\tau_1(X)$ in a preferred source is a minimizer of a population risk functional, with double-robustness: consistency holds if either the propensity or main-effect model is correctly specified in any source, and the method admits flexible weighting schemes derived from semiparametric or information-theoretic considerations (Li et al., 2022).

Causal-information-aware weighting further optimizes for variance reduction and efficiency: each source's contribution is scaled by a local density ratio and an information term inversely proportional to the semiparametric variance bound for CATE estimation. This ensures samples with greater alignment and higher information content are upweighted, with practical adaptive weight computation via modern ML or cross-fitting. Weighted multi-source direct learners (WMDL) outperform unweighted or single-source baselines across heterogeneous, high-dimensional data (Li et al., 2022).

5. Uncertainty Quantification: Bayesian Interventional Mean Processes

The quantification of uncertainty in fusion-based causal effect estimation leverages kernel mean embeddings and Gaussian process (GP) frameworks (Chau et al., 2021). The BayesIMP framework embeds both conditional and interventional distributions as elements in RKHS, fitting vector-valued GPs (BayesCME) to model conditional mean embeddings. Probabilistic integration combines two independent GP posteriors—one for the embedding $\mu_{Y|\mathrm{do}(X)=x}$, one for the outcome regression $f(y)=\mathbb{E}[T|Y=y]$—to yield a full posterior for the target $$g(x) = \langle f, \mu_{Y|\mathrm{do}(X)=x}\rangle_{\mathcal{H}}$$. Closed-form formulas for the posterior mean and covariance enable principled credible intervals and superior warm-start performance in Bayesian optimization of causal effects (Chau et al., 2021).

6. High-Dimensional and Algorithmic Considerations

In large causal graphs, preprocessing steps such as pruning (removing non-ancestor or irrelevant variables) and clustering (collapsing "transit clusters" that act equivalently towards the rest of the model) can radically reduce computational complexity without altering identifiability of the target effect (Tabell et al., 21 May 2025). Theorems guarantee that the do-calculus operations and final identifying functionals lift correctly from the reduced graph to the original, provided explicit compatibility and invariance conditions hold. Empirical gains can be orders of magnitude in runtime and search space.

Algorithmically, compositional approaches exploit advances in tractable probabilistic circuits (PCs) and md-vtrees to maintain and propagate "determinism" through sum-product representations, enabling polytime evaluation of complex composed queries including do-calculus adjustment, backdoor, and frontdoor estimands (Wang et al., 2023).

7. Empirical Performance and Application Domains

Simulations and real-world evaluations confirm that probabilistic causal fusion methods yield:

• Uniform reductions in mean squared error for dose-response estimation under source heterogeneity, even with non-smooth or nonparametric function classes (30–70% MSE improvement in low-$n$ settings) (Lim et al., 21 Oct 2025).
• Consistent minimax improvements in variance and coverage for joint causal effect vectors by combining constraints from randomization, IVs, and parental adjustments; fusion-based boosting methods achieve exact support recovery and lower risk than any single source (Gimenez et al., 2021).
• In spatiotemporal settings, fusion of spatially-aware causal inference (DiD, SAR models) and deep sequence modeling (RNNs, parametric output heads) achieves calibrated, probabilistic forecasts outperforming purely time-series or ML-based models especially under spatial spillover and intervention scenarios (Yang et al., 11 Jun 2025).
• Inference robustness to simultaneous unmeasured confounding and lack of exchangeability: partial identification methods yield meaningful bounds and breakdown frontiers delineating how much bias can be tolerated before causal claims reverse (Lanners et al., 30 May 2025).

Domains of application include health informatics (dose-response, trial+registry fusion), public policy (RCT, observational, and census integration), epidemiology (epidemics with regional interventions), and economics (meta-analysis, transportability).

---

In summary, probabilistic causal fusion unifies structural, algorithmic, and empirical advances in causal inference for multi-source, heterogeneous, and partially aligned data environments. Its toolkit of orthogonal empirical risk minimization, information-aware weighting, uncertainty quantification, graphical identification, and tractable computation enables sound and efficient estimation of causal effects under realistic, non-ideal conditions (Lim et al., 21 Oct 2025, Lee et al., 9 Apr 2024, Li et al., 2022, Gimenez et al., 2021, Chau et al., 2021, Lanners et al., 30 May 2025, Tabell et al., 21 May 2025, Yang et al., 11 Jun 2025, Wang et al., 2023).