📊 statistics

Causal-ICM: A Data Fusion Framework For Heterogeneous Treatment Effect Estimation With Multi-Task Gaussian Processes

The paper proposes Causal-ICM, a novel Bayesian nonparametric framework utilizing multi-task Gaussian processes to effectively fuse randomized controlled trial and observational data for robust heterogeneous treatment effect estimation with principled uncertainty quantification.

Original authors: Evangelos Dimitriou, Edwin Fong, Jens Magelund Tarp, Karla Diaz-Ordaz, Brieuc Lehmann

Published 2026-04-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Evangelos Dimitriou, Edwin Fong, Jens Magelund Tarp, Karla Diaz-Ordaz, Brieuc Lehmann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Perfect Lab" vs. The "Real World"

Imagine you want to know if a new medicine works. You have two sources of information:

The "Perfect Lab" (Randomized Controlled Trial or RCT): This is like a high-security, sterile laboratory. The scientists are very strict. They only let in healthy, young people with no other illnesses. Because they control everything perfectly, they know for a fact that the medicine caused the result. However, because the rules are so strict, the lab doesn't represent the messy, diverse real world. If you give this medicine to an elderly person with diabetes, the lab results might not apply to them.
The "Real World" (Observational Study): This is like a busy, chaotic city street. You look at thousands of people of all ages, backgrounds, and health conditions. This data is huge and very representative of the real world. But, it's messy. Maybe the sick people got the medicine because they were sicker, not because the medicine was good. There are hidden factors (confounders) messing up the data, making it hard to tell what actually caused the result.

The Dilemma:

The Lab is accurate but too small and narrow.
The City is big and diverse but messy and biased.

Most scientists struggle to combine these two. If they just mix them together, the huge amount of "messy" city data often drowns out the "accurate" lab data, leading to wrong conclusions. If they ignore the city data, they miss out on how the medicine works for regular people.

The Solution: Causal-ICM (The "Smart Translator")

The authors of this paper created a new tool called Causal-ICM. Think of it as a Smart Translator or a Bridge Builder between the Lab and the City.

Here is how it works, using a few metaphors:

1. The "Two-Task" Dance (Multi-Task Learning)

Imagine the Lab and the City are two dancers.

Dancer A (Lab) is dancing perfectly but in a small room.
Dancer B (City) is dancing wildly in a huge stadium, but sometimes trips over their own feet (bias).

Old methods tried to force them to dance the exact same steps. Causal-ICM says, "Let's watch them dance together." It uses a mathematical technique called Multi-Task Gaussian Processes. This is like a choreographer who watches both dancers and learns the rhythm of the Lab (the truth) while using the energy of the City (the volume) to fill in the gaps.

2. The "Volume Knob" (The Parameter $\rho$ )

This is the most clever part. The system has a special Volume Knob (called $\rho$ ).

If you turn the knob up to 100%, the system listens entirely to the City. This is dangerous because the City is messy.
If you turn the knob down to 0%, the system ignores the City and only listens to the Lab. This is safe but misses out on the big picture.

Causal-ICM's Superpower: It automatically finds the perfect middle setting. It listens to the City enough to learn about different types of people, but it keeps the Lab's "truth" loud enough so that the City's messiness doesn't drown it out. It prevents the "Big Messy Data" from tricking the system into being overconfident.

3. The "Safety Net" (Uncertainty Quantification)

Imagine you are trying to guess the weather.

The Lab says: "It will rain at 2 PM." (Very sure, but only for Tuesday).
The City says: "It might rain, or snow, or shine!" (Very unsure, but covers every day).

If you just average them, you might get a wrong answer with a false sense of confidence. Causal-ICM is different. If it has to guess about a day the Lab never tested (extrapolation), it doesn't just give you a number; it puts up a Safety Net. It says, "I think it will rain, but because I'm guessing outside the Lab's data, I'm only 80% sure."

It mathematically guarantees that even if the City data is huge, the system will never become too confident if the data is biased. It keeps a healthy dose of "I'm not 100% sure" in its answers.

How They Tested It

The authors didn't just talk about it; they put it to the test in two ways:

The Simulation (The Video Game): They created fake worlds where they knew the "true" answer. They pitted Causal-ICM against other famous methods.
- Result: Causal-ICM won. It was better at guessing the truth, especially in the "messy" parts of the data where other methods failed or became overconfident.
The Real World (The STAR Study): They used real data from a famous 1980s study about class sizes in schools. They tried to figure out how class size affects student test scores.
- Result: Causal-ICM performed just as well as the best existing methods, but with the added benefit of knowing exactly how uncertain it was about its predictions.

Why This Matters

In the real world, we rarely have perfect data. We have to make decisions about medicine, policy, and economics based on imperfect information.

Causal-ICM is like a wise advisor. It takes the strict, reliable facts from controlled experiments and combines them with the broad, real-world experience of observational data. But unlike a naive advisor who gets swayed by the loudest voice (the biggest dataset), Causal-ICM knows how to listen carefully, keep its guard up against bias, and always tell you how sure it is about its advice.

This helps doctors and policymakers make better, safer decisions for everyone, not just the people who fit perfectly into a clinical trial.

1. Problem Statement

The paper addresses the challenge of estimating Heterogeneous Treatment Effects (HTE)—specifically the Conditional Average Treatment Effect (CATE)—by bridging the gap between internal validity and external validity.

Randomized Controlled Trials (RCTs): Provide high internal validity (unbiased estimates due to randomization) but often suffer from limited sample sizes, strict eligibility criteria, and poor representativeness of the target population (low external validity).
Observational Studies: Offer large sample sizes and better representativeness (high external validity) but are susceptible to unobserved confounding, leading to biased treatment effect estimates.

Existing data fusion methods often rely on strong parametric assumptions (e.g., linear confounding) or fail to provide principled uncertainty quantification, particularly when extrapolating beyond the RCT support. There is a need for a flexible, non-parametric approach that can integrate these data sources while safeguarding against the bias introduced by the observational data.

2. Methodology: Causal-ICM

The authors propose Causal-ICM, a Bayesian nonparametric framework based on Multi-Task Gaussian Processes (GP) using Intrinsic Coregionalization Models (ICM).

Core Framework

Multi-Task Formulation: The method treats the outcome regression functions for the RCT ( $f_e$ $f_{e}$ ) and the observational study ( $f_o$ $f_{o}$ ) as two distinct but related tasks.
- $f_e(x) = E[Y | x, S=e, A=1]$
- $f_o(x) = E[Y | x, S=o, A=1]$
ICM Structure: The model assumes that both tasks are linear combinations of independent latent functions ( $u_1, u_2$ $u_{1}, u_{2}$ ) drawn from the same scalar GP prior.
- $f_e(x) = \alpha^e_1 u_1(x) + \alpha^e_2 u_2(x)$
- $f_o(x) = \alpha^o_1 u_1(x) + \alpha^o_2 u_2(x)$
Parameterization for Causal Inference: The authors introduce a specific, interpretable parameterization to control information flow:
$\begin{bmatrix} \alpha^e_1 \\ \alpha^e_2 \\ \alpha^o_1 \\ \alpha^o_2 \end{bmatrix} = \begin{bmatrix} 1 \\ 0 \\ \rho \\ \sqrt{1-\rho^2} \end{bmatrix}$
This yields:
- $f_e(x) = u_1(x)$ (The RCT surface is modeled directly).
- $f_o(x) = \rho f_e(x) + \sqrt{1-\rho^2} u_2(x)$ (The observational surface is a scaled version of the RCT surface plus a task-specific noise/confounding term).

The Role of $\rho$ (Borrowing Parameter)

$\rho \in (0, 1)$ controls the degree of "borrowing" from the observational data.
$\rho \to 1$ : Maximum borrowing (observational data dominates).
$\rho \to 0$ : No borrowing (RCT data dominates).
Variance Bound (Proposition 1): A key theoretical contribution is the proof that the posterior variance of the RCT estimate, $V^e(x^*)$ $V^{e} (x^{*})$ , is lower-bounded by $(1-\rho^2)V^e_{D_e}(x^*)$ $(1 - ρ^{2}) V_{D_{e}}^{e} (x^{*})$ , where $V^e_{D_e}$ $V_{D_{e}}^{e}$ is the variance using only RCT data.
- Implication: Even as the observational sample size ( $n_o$ ) approaches infinity, the uncertainty in the estimate cannot shrink arbitrarily. This prevents the model from becoming overconfident in the presence of unobserved confounding.

Tuning and Estimation

Data-Adaptive Selection: Instead of maximizing marginal likelihood (which might overfit the large observational dataset), $\rho$ is selected via 5-fold cross-validation on a held-out RCT subset. The objective minimizes the Root Mean Squared Error (RMSE) with weights that upweight RCT patients who resemble the observational population (to optimize extrapolation).
CATE Estimation: The framework uses a T-learner approach, fitting separate ICMs for the treatment ( $A=1$ ) and control ( $A=0$ ) arms to derive the posterior distribution of $\tau(x) = f_e^1(x) - f_e^0(x)$ .

3. Key Contributions

Novel Bayesian Nonparametric Model: Introduces Causal-ICM, a multi-task GP framework specifically designed for causal data fusion that avoids strong parametric assumptions about the bias function or CATE structure.
Theoretical Safeguard Against Bias: Proves that the model limits the information flow from the observational dataset. The posterior variance is bounded, ensuring that large observational samples do not artificially narrow uncertainty intervals around biased estimates.
Data-Adaptive Hyperparameter Tuning: Proposes a cross-validation procedure to select the borrowing parameter $\rho$ , balancing precision gains against bias risks.
Robust Extrapolation: Demonstrates the ability to extrapolate treatment effects to regions of the covariate space not covered by the RCT, leveraging the broader support of observational data while maintaining valid uncertainty quantification.

4. Results

The authors evaluated Causal-ICM through extensive simulations and a real-world dataset (Tennessee STAR study).

Simulation Studies:
- Linear vs. Non-linear: In scenarios with non-linear CATE and non-linear confounding, Causal-ICM significantly outperformed state-of-the-art methods (e.g., Integrative HTE, Elastic HTE, Experimental Grounding) which often rely on linear assumptions.
- Uncertainty Quantification: Causal-ICM achieved near-nominal (95%) coverage for credible intervals. In contrast, observational-only methods produced overconfident intervals (near 0% coverage) as sample size increased, failing to account for bias.
- Scalability: The method remained robust as the observational sample size grew, whereas methods like Integrative HTE degraded in performance under model misspecification.
Real-World Data (Tennessee STAR):
- Applied to estimate the effect of class size on student test scores.
- Causal-ICM achieved competitive RMSE (6.29), comparable to the best-performing baseline (Experimental Grounding with Random Forests at 6.27) and significantly better than the observational-only T-learner (7.47) or Integrative HTE (11.84).
Extrapolation: The model successfully provided accurate estimates and inflated uncertainty in regions outside the RCT support, a critical feature for generalizability.

5. Significance

Causal-ICM represents a significant advancement in causal inference and data fusion:

Principled Uncertainty: It solves the critical problem of "overconfidence" in data fusion, where large observational datasets can mask bias. By theoretically bounding the variance reduction, it ensures that decision-makers are aware of the uncertainty introduced by unobserved confounding.
Flexibility: By using Gaussian Processes, it captures complex, non-linear relationships without requiring the researcher to specify the functional form of the treatment effect or the confounding mechanism.
Decision Support: The ability to provide reliable HTE estimates with valid uncertainty intervals across the full target population supports more personalized medical and policy decisions, particularly for subgroups underrepresented in RCTs.
Future Directions: The framework opens avenues for handling non-Gaussian outcomes, multiple data sources, and hierarchical confounders, positioning multi-task GPs as a powerful tool for next-generation evidence synthesis.