📊 statistics

Interventional Processes for Causal Uncertainty Quantification

This paper introduces a Gaussian process-based framework that leverages reproducing kernel Hilbert space representations to provide reliable, tractable, and calibrated uncertainty quantification for interventional causal functions, addressing the challenges of estimating entire functions rather than scalar values.

Original authors: Hugh Dance, Peter Orbanz, Arthur Gretton

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hugh Dance, Peter Orbanz, Arthur Gretton

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: Guessing the "What If" with Confidence

Imagine you are a doctor trying to decide if a new medicine works. You can't just look at people who took the medicine and say, "It worked!" because maybe those people were already healthier. You need to know: If we force a specific person to take the medicine, what happens?

In statistics, this is called a causal effect. The hard part isn't just guessing the answer (the "point estimate"); it's knowing how sure you can be about that answer.

Scalar vs. Function: Usually, people ask, "Does the drug work on average?" (A single number). But in real life, the answer changes depending on the patient's age, weight, or genetics. The answer is a whole function (a curve showing how the drug works for everyone).
The Risk: If you only give a single number without a "confidence range," you might make a dangerous decision. If you give a confidence range that is too narrow (overconfident), you might think a drug is safe when it's actually risky.

The Old Way: Trying to Fit a Square Peg in a Round Hole

Previous methods tried to use a tool called a Gaussian Process (GP) to draw these confidence curves. Think of a GP as a flexible rubber sheet that tries to fit the data.

However, the math behind "causal functions" is very strict. It lives in a special mathematical space called an RKHS (Reproducing Kernel Hilbert Space).

The Problem: Trying to force a standard rubber sheet (GP) to fit inside this strict, narrow space is like trying to push a square peg into a round hole.
The Result: The old methods (like "BayesIMP") tried to force it anyway. This caused two major glitches:
1. Underfitting: The rubber sheet got too stiff and couldn't bend enough to follow the true shape of the data. It missed the peaks and valleys.
2. Variance Collapse: When you asked the model about a patient type it hadn't seen before (outside the training data), the rubber sheet suddenly snapped flat. The model became overconfident, saying, "I know exactly what happens here!" when it actually knew nothing.

The New Solution: IMPspec (The Flexible Bridge)

The authors, Hugh Dance, Peter Orbanz, and Arthur Gretton, built a new method called IMPspec (Interventional Mean Process with Spectral representation).

Here is how they fixed the problems, using a simple analogy:

1. Relaxing the Rules (The "Loose Rope" Analogy)

Instead of forcing the rubber sheet to stay strictly inside the narrow "causal" room, they expanded the room.

They realized that if they let the rubber sheet be slightly "looser" (allowing it to exist in a broader space), they could still pull the final answer back into the strict causal room.
The Magic Trick: They used a technique called Spectral Representation. Imagine the complex causal function is a song. Instead of trying to write the whole song at once, they broke it down into individual musical notes (frequencies). They put a standard, easy-to-use rubber sheet on each note.
The Benefit: Because they are working with individual notes, they can use standard, reliable tools. When they put the notes back together, the final song is perfect, and the math stays simple.

2. The "Calibration" Step (The "Thermostat" Analogy)

Even with the new method, the confidence intervals (the "uncertainty bands") might be slightly off—maybe a little too tight or a little too loose.

The authors added a Calibration step. Think of this like a thermostat.
They run a simulation (a "bootstrap") where they pretend to re-run the experiment many times. They check: "If we say we are 95% sure, does the truth actually fall inside our range 95% of the time?"
If the answer is "No, we are only right 80% of the time," the thermostat adjusts the "knob" (the spectral measure) until the confidence intervals are perfectly calibrated.

What They Found (The Results)

The paper tested this new method in three ways:

Toy Examples: In simple, made-up scenarios, IMPspec was much better at drawing the curve than the old methods. The old methods (BayesIMP) were "underfitting" (missing the shape) and "collapsing" (getting overconfident in new areas). IMPspec stayed flexible and honest about its uncertainty.
Synthetic Benchmarks: They created complex fake worlds with hidden variables. IMPspec again produced the most accurate curves and the most reliable confidence intervals.
Real Data (401(k) and Healthcare):
- They used a real dataset about 401(k) retirement plans to see how eligibility affects wealth. IMPspec showed that the effect grows with income, and the uncertainty bands were reasonable.
- They used it for Causal Bayesian Optimization. Imagine you are trying to find the perfect dose of a drug to minimize cancer volume, but you can only test a few doses. IMPspec acted as a better "guide," helping the search find the best dose faster and more safely than other methods because it knew exactly where it was unsure.

The Bottom Line

This paper introduces a new way to calculate how sure we are about the effects of interventions (like drugs or policies).

Old Way: Tried to force a square peg into a round hole, leading to stiff, inaccurate models that got dangerously overconfident.
New Way (IMPspec): Breaks the problem into smaller, manageable pieces (spectral representation), uses standard tools to solve them, and then calibrates the final result like a thermostat.
Result: It gives us a "rubber sheet" that is flexible enough to fit the data perfectly but honest enough to admit when it doesn't know the answer, making high-stakes decisions safer.

Technical Summary: Interventional Processes for Causal Uncertainty Quantification

Problem Statement

Reliable uncertainty quantification (UQ) for causal effects is critical in high-stakes domains like healthcare and policy. While frequentist methods provide valid confidence intervals for scalar average treatment effects (ATEs), extending these guarantees to function-valued causal targets (e.g., conditional average treatment effects (CATEs) or continuous dose-response curves) remains challenging.

Existing approaches face significant hurdles:

Frequentist Limitations: Standard debiased estimators rely on Riesz representers, which may not exist for function-valued targets without strong parametric or smoothness assumptions. These methods are also sensitive to weak overlap in finite samples.
Bayesian Limitations: Placing priors on the conditional laws that identify interventional functions often leads to computationally intractable posterior inference.
Gaussian Process (GP) Limitations: While GPs are standard for functional UQ, constructing priors specifically for functions within a Reproducing Kernel Hilbert Space (RKHS) is non-trivial. Previous attempts (e.g., BayesIMP) using bespoke covariance constructions suffer from pathologies, including underfitting due to non-stationary kernels and variance collapse outside the training support.

Methodology: IMPSPEC

The authors propose IMPSPEC (Interventional Mean Process via Spectral representation), a GP-based framework that quantifies uncertainty over causal functions by leveraging their RKHS representations. The core innovation lies in relaxing strict RKHS constraints to enable the use of standard GP machinery while preserving the theoretical guarantees of the underlying kernel estimators.

1. Relaxing RKHS Constraints via Spectral Representation

Instead of enforcing the original RKHS constraints directly on the regression objects, the method expands the function classes:

For the Outcome Regression ( $f$ ): The authors place a simple linear-kernel GP prior directly on the functional $f$ , rather than constructing a bespoke prior to enforce boundedness. This enlarges the function class but allows for closed-form posterior updates.
For the Conditional Embedding ( $\mu$ ): Since $\mu(z) = E[\psi_V(V) | Z=z]$ is an infinite-dimensional RKHS-valued function, standard GPs do not apply. The authors utilize Mercer's theorem to decompose the feature map $\psi_V$ into a spectral series of eigenfunctions. This reduces $\mu$ to a sequence of scalar-valued conditional expectations ( $\mu_i$ ), on which standard scalar GP priors can be placed independently.

2. The Interventional Process

The causal function $\gamma(w, z)$ is represented as an inner product in the RKHS: $\gamma(w, z) = f(\psi_W(w) \otimes \mu(z))$ . By combining the GP priors on $f$ and the spectral coordinates $\mu_i$ , the authors define an Interventional Process.

Posterior Mean: The construction ensures the posterior mean of $\gamma$ exactly recovers the kernel causal estimator proposed by Singh et al. (2024), thereby targeting uncertainty around this established point estimate.
Posterior Variance: The authors derive closed-form expressions for the posterior variance, which decompose into three terms ( $S_1, S_2, S_3$ ) accounting for uncertainty in $f$ , uncertainty in $\mu$ , and the interaction between them.

3. Hyperparameter Training and Calibration

Training: Hyperparameters (kernel lengthscales, noise variances) are optimized by maximizing the marginal likelihood. For the spectral components, the authors aggregate coordinate-wise likelihoods into a single tractable weighted objective.
Calibration: To address potential coverage errors arising from surrogate likelihoods and moment-matched Gaussian approximations, the authors introduce a calibration procedure. This tunes the spectral measure $\nu$ (a free parameter) to minimize the discrepancy between nominal credible levels and empirical coverage, using a bootstrap-based loss function.

Key Contributions

Novel GP Framework: Introduction of a GP-based approach for UQ of interventional functions that avoids the pathologies of previous RKHS-constrained GP constructions.
Closed-Form Inference: Derivation of closed-form posterior moments (mean and variance) for the interventional process, enabling tractable training and inference without Monte Carlo sampling of the posterior.
Consistency with Point Estimators: The posterior mean exactly recovers the kernel causal estimator of Singh et al. (2024), ensuring the UQ is centered on a state-of-the-art point estimate.
Calibration Procedure: A practical algorithm for posterior coverage calibration via spectral measure tuning, improving the reliability of credible intervals.
Empirical Validation: Demonstration of improved uncertainty quantification across synthetic benchmarks, causal Bayesian optimization tasks, and a large-scale real-world dataset (401(k)), while maintaining competitive performance in effect estimation.

Experimental Results

The authors evaluate IMPSPEC against several baselines, including BayesIMP, Sampling GPs, Orthogonal Random Forests (ORF), and doubly robust estimators.

Toy Example: IMPSPEC achieved the lowest Root Mean Squared Error (RMSE) and calibration error. Notably, BayesIMP systematically underfit the causal function, attributed to the non-stationarity of its kernel construction.
Synthetic Benchmark: IMPSPEC outperformed all methods in RMSE, calibration error, and interval score. Crucially, unlike BayesIMP, IMPSPEC did not suffer from variance collapse outside the training support, maintaining reliable uncertainty estimates in out-of-distribution regions.
Causal Bayesian Optimization (CBO): In tasks aiming to find optimal interventions, IMPSPEC converged fastest and achieved the lowest cumulative regret. In one instance, it was the only method to find the approximate optimum within the iteration limit, highlighting the utility of its uncertainty estimates for decision-making.
Real-World Data (401(k)): Applied to a dataset of 9,915 observations, IMPSPEC produced plausible uncertainty-aware estimates of the income-dependent CATE of 401(k) eligibility on financial assets. The method demonstrated computational feasibility, running in approximately one minute on a GPU with manageable memory usage.

Significance and Claims

The paper claims that IMPSPEC provides a tractable and reliable solution for uncertainty quantification in causal inference where the target is a function rather than a scalar. By bridging the gap between non-parametric kernel estimators and Gaussian processes, the method offers:

Robustness: Avoidance of variance collapse and underfitting seen in prior RKHS-constrained GP methods.
Practicality: Closed-form inference and efficient hyperparameter tuning suitable for large-scale datasets.
Reliability: Improved calibration of credible intervals, which is essential for high-stakes decision-making where overconfidence can be dangerous.

The authors emphasize that while the method improves UQ, its reliability in high-stakes applications ultimately depends on the appropriateness of the underlying identification assumptions (e.g., the validity of the causal graph) and the quality of the calibration procedure. They position IMPSPEC as a tool to be used alongside domain expertise and sensitivity analysis.