FADO: Learning Causal Explanations with Prior-Fitted Networks

Seth Flaxman, Jess Carr, Annabel Jakob, Marek Masiak, Dino SejdinovicarXiv 2026 · arXiv

TL;DR

FADO (Feature-wise Amortized Do-Operators) is a novel causal explainability method that uses Prior-Fitted Networks to estimate the impact of intervening on features directly from observational data. By training a transformer on millions of synthetic causal models, FADO learns to distinguish between features that are merely predictive (correlated) and those that are truly causal, offering a faster and more robust alternative to traditional causal discovery and predictive attribution methods like SHAP.

What Problem Does This Paper Address?

Most popular machine learning explainability tools, such as SHAP or permutation importance, focus on predictive attribution: they answer the question, "Does this feature help the model predict the outcome?" While useful, this approach captures correlation rather than causation. In scientific and high-stakes domains (like healthcare or finance), we often need to answer a different, interventional question: "If I change this feature, will the outcome actually change?"

Standard predictive methods often fail in common causal scenarios. For example, they might assign high importance to a "collider" (a variable caused by both the feature and the outcome) or a "descendant" of the target, even though neither has a causal effect on the outcome. Conversely, they might miss "upstream" causal variables whose effects are fully mediated by other features. FADO aims to bridge this gap by providing causal attribution directly from observational tables.

Methods

The authors introduce FADO, which stands for Feature-wise Amortized Do-Operators. The core idea is "learning to explain" through amortized Bayesian inference.

  1. Target Functional: FADO targets the standardized feature intervention effect ($\tau_i$), defined as the difference in the expected outcome when a feature is intervened upon (set to +1 vs. -1), standardized by the outcome's standard deviation.
  2. Prior-Fitted Networks (PFNs): Instead of fitting a model to a specific dataset and then explaining it, the authors train a single large Transformer model on a massive variety of synthetic datasets generated by Structural Causal Models (SCMs). Because these are synthetic, the ground-truth causal effects are known during training.
  3. Axial Transformer Architecture: To handle tabular data, FADO uses an Axial Transformer that alternates attention along two axes:
  4. Feature axis attention: Cells within a row attend to each other to understand relationships between variables.
  5. Sample axis attention: Cells within a column attend to each other to aggregate information across observations.
  6. Amortization: At test time, FADO requires only a single forward pass through the network to predict the causal importance of every feature in a new, unseen observational dataset.
Figure 2: FADO architecture. Cell values are projected to d dimensions and summed with a learned type embedding. Each of the 6 layers applies feature axis attention within rows, then sample axis attention within columns. The cell embeddings are averaged over samples, and a linear head at the feature positions emits ˆτ ∈Rp.

Datasets

The authors evaluate FADO across a diverse battery of benchmarks:

  • Synthetic SCMs: 11 different prior families including linear, polynomial, MLP-based, and models with hidden confounders.
  • Causal Motifs: Five prototypical small-scale causal structures (Direct Cause, Confounder, Mediator, Descendant, Collider) designed to test specific failure modes of predictive attribution.
  • Semi-Synthetic (ACIC-2016): Real-world covariates from a social science study, but with simulated outcomes and "decoy" variables (noise, proxies, and outcome leaks).
  • Real RCTs: The Hillstrom e-mail marketing dataset and "RCT Bench" (a collection of 124 real-world trials).
  • Physical Systems: Data from the "Causal Chambers," a physical light tunnel device with measured intervention effects.

Key Insights and Ideas

  • Predictive vs. Causal Disagreement: The paper beautifully illustrates how SHAP and FADO can disagree. As shown in Figure 1, features like "colliders" or "descendants" are highly predictive but not causal. FADO correctly assigns them near-zero importance, whereas SHAP ranks them highly.
Comparison of predictive (SHAP) vs causal (FADO) importance
  • Handling "Outcome Leaks": In real-world data science, "leaks" (variables that are functions of the outcome) are common and fool predictive models. FADO is shown to be significantly more robust to these leaks than standard methods.
  • Amortization Efficiency: Because the "heavy lifting" is done during pre-training, FADO is incredibly fast at inference time—roughly 1,100x faster than Causal Forests for a table with 13 features.
  • Uncertainty as a Guide: By using bootstrap resamples, FADO can provide uncertainty estimates. The authors found that FADO’s self-reported uncertainty is a strong predictor of its actual error, allowing users to trust the model more when it is "certain."

Key Results

  • State-of-the-Art Performance: FADO outperformed every other baseline (including LiNGAM, NOTEARS, and Causal Forests) across almost every synthetic regime tested, particularly in non-linear settings.
  • Robustness to Decoys: In semi-synthetic tests with real covariates, FADO was the best at zeroing out "pure noise" and "outcome leak" variables.
  • Real-World Generalization: On 124 real-world Randomized Controlled Trials (RCTs), FADO achieved a 0.85 correlation with the ground-truth treatment effects, despite not being specifically designed for single-treatment estimation.
  • Scaling: While the main model was trained for up to 15 features, the authors successfully extended the approach to tables with up to 64 features.

Limitations and Future Work

  • Identifiability: FADO is a Bayesian estimator. If a causal effect is fundamentally unidentifiable from observational data (e.g., in a perfectly linear Gaussian system where directionality is ambiguous), FADO will predict the prior-weighted average of possible effects rather than the "true" one.
  • Point Estimation: The authors note that while FADO is excellent for ranking feature importance, it is not yet a replacement for high-precision point estimation of specific treatment effects.
  • Symmetric Mechanisms: The current (-1, +1) contrast might miss causal effects that are perfectly symmetric around zero (e.g., $Y = X^2$). The authors suggest widening the "head" of the model to a $K$-anchor curve readout as a future improvement.
  • Tabular Only: The current work is focused strictly on tabular data; extending this "learning to explain" philosophy to images or text remains an open challenge.

Summary updated

Abstract

In regression and classification tasks with tabular datasets, most of the literature on explanation in ML (e.g. SHAP, permutation importance) has focused on predictive attribution: does feature $X_i$ help predict outcome $Y$? At its heart, predictive attribution captures correlation, not causation. We instead ask a causal question: would manipulating feature $X_i$ change outcome $Y$? We define a natural quantity to answer this question, the feature intervention effect: $Δ_i = E[Y\mid do(X_i=+1)]-E[Y\mid do(X_i=-1)]$. In general, estimating this quantity from observational data alone would entail strong causal assumptions. We propose FADO (Feature-wise Amortized Do-Operators), an amortized inference method using prior-fitted networks trained on synthetic tabular datasets generated by structural causal models. During training, the synthetic data-generating process is known, so we can compute ground-truth intervention effects $Δ_i$, and train a model to predict them at test time from a purely observational tabular dataset. As we demonstrate through extensive experiments, our new causal explainability method is easy to train and provides better explanations under a range of settings for which standard predictive attribution fails, including proxy variables, confounders, colliders, and placebos. We evaluate it on real, synthetic and semi-synthetic datasets, with and without hidden confounders, and compare it to existing predictive attribution methods and state-of-the-art causal inference methods.
BibTeX
@misc{flaxman2026fado,
  author = {Seth Flaxman and Jess Carr and Annabel Jakob and Marek Masiak and Dino Sejdinovic},
  title = {{FADO: Learning Causal Explanations with Prior-Fitted Networks}},
  year = {2026},
  eprint = {2611.15607},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15607},
}
Figure