Debiased Front-Door Learners for Heterogeneous Effects

Conference: ICLR 2026
Code: https://github.com/yonghanjung/FD-CATE
Area: Causal Inference / Heterogeneous Treatment Effects
Keywords: Front-door adjustment, Heterogeneous treatment effects, Debiased learning, quasi-oracle rates, DR-Learner, R-Learner

TL;DR

This paper transplants the mature DR-Learner and R-Learner from back-door settings to front-door identification scenarios. It proposes two debiased estimators, FD-DR-Learner and FD-R-Learner, ensuring that the conditional front-door effect \(\tau(C)\) achieves quasi-oracle rates even when nuisance functions converge at a slow rate of \(n^{-1/4}\).

Background & Motivation

Background: Causal inference from observational data faces the significant challenge of unobserved confounding—where the treatment \(X\) is interfered with by latent variables \(U\) that affect both \(X\) and the outcome \(Y\). In such cases, \(E[Y|X=1]-E[Y|X=0]\) is biased. Pearl's front-door criterion provides a solution: identifying an observable mediator \(Z\) that transmits the effect of \(X\) to \(Y\) (e.g., "Active seatbelt law \(X\) → Seatbelt usage rate \(Z\) → Occupant fatality \(Y\)"). As long as \(Z\) itself is not confounded by \(U\), the causal effect can be identified by bypassing the latent confounding between \(X\) and \(Y\).

Limitations of Prior Work: Although debiased estimation for the front-door direction has been progressed (Fulcher 2019, Guo 2023, Jung 2024, etc.), nearly all focus on estimating the Average Treatment Effect (ATE). However, platforms and policymakers often require the individualized conditional front-door effect \(\tau(C)\). On another track, while deep estimators for heterogeneous effects exist (LobsterNet by Xu & Gretton 2022, Chen 2025), they lack debiasedness—the estimation fails if the nuisance functions are not accurately fitted. In other words, "front-door + heterogeneity + debiasedness" have never been simultaneously satisfied (see comparison in Table 1c of the original paper).

Key Challenge: The strength of DR/R-Learners under the back-door setting stems from the debiasedness brought by Neyman orthogonalization, where slow convergence of nuisance functions does not hinder the target. However, the front-door estimator structure is more complex (involving three sets of nuisance functions \(m, e, q\) and their density ratio combinations), making it impossible to directly copy the pseudo-outcome construction from the back-door setting.

Goal: Construct front-door versions of pseudo-outcomes and orthogonal losses, allowing "arbitrary off-the-shelf ML models + slow nuisances" to converge rapidly to \(\tau(C)\). Core Idea: (1) FD Pseudo-Outcome (FDPO) formulates the front-door effect as a regressible quantity where nuisance errors only appear as second-order terms; (2) Partial Differential Linear Reparameterization decomposes the front-door effect into two standard back-door R-Learner sub-problems \(b(C)\) (\(X \to Z\)) and \(g(XC)\) (\(Z \to Y\)), using pseudo-g to decouple the error of the combined term \(\gamma_g\) from \(\hat{e}_X\).

Method

Overall Architecture

Two learners tackle the same goal \(\tau(C)=\sum_{z,x}\{q(z|1C)-q(z|0C)\}e_x(C)m(zxC)\) via different paths: FD-DR-Learner follows a "single pseudo-outcome regression" path—constructing a pseudo-outcome whose conditional mean exactly equals \(\tau_{\bar x}(C)\) and regressing it onto \(C\); FD-R-Learner follows a "decompose-and-combine" path—rewriting the data generation process into two partial differential linear models, learning the path coefficients \(b\) and \(g\) using off-the-shelf back-door R-Learners, and then synthesizing them. Both rely on Neyman orthogonal structures to achieve second-order dependence on nuisance errors.

flowchart TD
    A[Observed Data V= C,X,Z,Y <br/>Front-Door: X→Z→Y, U confounds X,Y] --> B{Two Debiased Paths}
    B --> C[FD-DR-Learner]
    B --> D[FD-R-Learner]
    C --> C1[Fit nuisance m, e, q]
    C1 --> C2[Construct FD Pseudo-Outcome φ_x̄<br/>Includes density ratios ξ, π + correction terms]
    C2 --> C3[Regress φ_1-φ_0 to C → τ̂_DR]
    D --> D1[Partial Differential Linear Reparameterization<br/>X→Z yields b, Z→Y yields g]
    D1 --> D2[Apply BD-R-Learner to learn b, g]
    D2 --> D3[pseudo-g decouples ê_X error → γ̂]
    D3 --> D4[τ̂_R = b̂·γ̂]

Key Designs

1. Front-Door Pseudo-Outcome (FDPO): Formulating the effect as a regressible quantity with second-order sensitivity to nuisance errors. The core of FD-DR is constructing a pseudo-outcome \(\varphi_{\bar x}(V;\eta)\) for each intervention value \(\bar x\). it consists of three parts: a residual \(Y-m(ZXC)\) weighted by the density ratio \(\xi_{\bar x}(ZXC)=q(Z|\bar xC)/q(Z|XC)\), a correction term \(r_{me}(ZC)-\nu_{meq}(XC)\) weighted by the inverse propensity \(\pi_{\bar x}(XC)=\mathbb{I}(X=\bar x)/e(X|C)\), and a direct term \(s_{mq\bar x}(XC)\):

\[\varphi_{\bar x}(V;\eta)=\xi_{\bar x}\{Y-m\}+\pi_{\bar x}\{r_{me}-\nu_{meq}\}+s_{mq\bar x}.\]

The elegance of this construction lies in two properties provided by Lemma 2: Consistency \(\tau_{\bar x}(C)=E[\varphi_{\bar x}(V;\eta)\mid C]\), meaning regressing \(\varphi_1-\varphi_0\) onto \(C\) directly yields \(\tau(C)\); and Double Robustness—when replacing the true values with estimates \(\hat\eta\), the bias \(E[\varphi_{\bar x}(V;\hat\eta)-\varphi_{\bar x}(V;\eta)]\) consists entirely of cross-products of nuisance errors (e.g., \(\{\hat m-m\}\{\xi-\hat\xi\}\)). Thus, if either \(\hat q\) is accurate or \((\hat m, \hat e)\) are accurate, the first-order bias is canceled. Theorem 1 establishes \(\|\hat\tau_{DR}-\tau\|_2^2\lesssim R_{DR}+\sum\|\hat m-m\|^2\|\hat\xi-\xi\|^2+\dots\), achieving quasi-oracle rates when all nuisances reach \(n^{-1/4}\).

2. Partial Differential Linear Reparameterization for Front-Door: Splitting one front-door problem into two back-door R-Learners. Instead of directly handling the complex front-door estimator, FD-R first proves that the front-door structure is equivalent to a set of layered partial differential linear models (Prop. 2): \(Z=a(C)+Xb(C)+\epsilon_Z\) describes \(X \to Z\), and \(Y=f(XC)+Zg(XC)+\epsilon_Y\) describes \(Z \to Y\). Since \(C\) satisfies the back-door criterion for \((X, Z)\) and \((X, C)\) for \((Z, Y)\), \(b(C)\) and \(g(XC)\) can be learned using standard BD-R-Learners, inheriting their debiasedness (slow nuisances do not impede convergence). Theorem 2 further proves that the heterogeneous front-door effect can be written as the product of these two path coefficients:

\[\tau(C)=b(C)\,\gamma_g(C),\qquad \gamma_g(C)=E[g(XC)\mid C].\]

This steps translates a "hard" front-door estimation into two "solved" sub-problems, with the added benefit that \(b\) and \(g\) serve as interpretable intermediate quantities of path intensities for \(X \to Z\) and \(Z \to Y\).

3. pseudo-g: Decoupling the error of the combined term \(\gamma_g\) from the propensity score \(\hat e_X\). After obtaining \(\hat b, \hat g\), one must estimate \(\gamma_g(C)=e_X(C)g(1C)+\{1-e_X(C)\}g(0C)\). A naive plug-in \(\hat\gamma_{plug}\) directly substitutes \(\hat e_X\), but its error \(\hat\gamma_{plug}-\gamma_g\) contains an \(\{g(1C)-g(0C)\}(\hat e_X-e_X)\) term—creating a bottleneck based on the propensity score's accuracy. This paper instead defines pseudo-g:

\[\zeta_{\tilde\eta_z}(XC)=\{1-e_X\}\tilde g(0C)+e_X\,\tilde g(1C)+\{X-e_X\}\{\tilde g(1C)-\tilde g(0C)\}.\]

Lemma 3 shows it satisfies \(E[\zeta_{\eta_z}\mid C]=\gamma_g(C)\), and the bias of \(\hat e_X\) in the error correction term is canceled—\(E[\zeta_{\hat\eta_z}\mid C]-\gamma_g\) contains only \(e_X\{\hat g(1C)-g(1C)\}+\{1-e_X\}\{\hat g(0C)-g(0C)\}\), which is purely determined by the error of \(\hat g\), which can be efficiently learned via BD-R-Learner. Finally, regressing \(\zeta\) onto \(C\) yields \(\hat\gamma\), returning \(\hat\tau_R(C)=\hat b(C)\hat\gamma(C)\). Theorem 3 shows its error is controlled by the quasi-oracle rate plus second/fourth-order nuisance error terms, similarly achieving quasi-oracle at \(n^{-1/4}\).

Key Experimental Results

Synthetic Experiments (True \(\tau(C)\) known, nuisances using XGBoost)

Comparing RMSE (mean ± 95% CI) of the plug-in baseline FD-PI, FD-DR, and FD-R across four regimes:

Experimental Setting	FD-PI (plug-in)	FD-DR	FD-R	Conclusion
(a) Varied Sample Size \(n\), no structural noise	High	Low	Low	Both debiased learners consistently outperform plug-in
(b) Nuisance restricted to \(n^{-1/4}\) slow rate	Very slow convergence	Reliable convergence	Reliable convergence	Validates debiasedness
(c) Injected noise \(\rho\epsilon,\ \rho\in[0,1]\)	Deteriorates sharply with \(\rho\)	Stable	Lower and more stable	FD-R is least sensitive to nuisance error
(d) Weak overlap (positivity approaches 0/1)	Severe degradation	Variance inflation (due to weights)	Most stable, leading throughout	FD-R is robust to weak overlap as it avoids density ratios

Real Case: State Seatbelt Laws and Fatality Rates (FARS)

State-year panel data, \(X\)=Active seatbelt law presence, \(Z\)=Seatbelt usage rate, \(Y\)=Occupant fatalities, \(C\)=Covariates:

• The distribution of \(\hat\tau\) estimated by both learners is negatively skewed (laws reduce fatalities), with overall means of approximately \(-0.047\) (FD-DR) and \(-0.046\) (FD-R).
• Concentration curves show that >95% of units saw a reduction in fatality rates under active laws, with only a few showing increases, aligning with the expected protective effect of seatbelt laws.
• SHAP attribution indicates that Age, Time of Day, and Driver status are the dominant features explaining effect heterogeneity.

Key Findings

• Debiasedness is empirically verifiable: When nuisances only reach the \(n^{-1/4}\) slow rate, the plug-in significantly lags, while FD-DR/FD-R still converge rapidly, confirming the theory.
• Operational division between FD-DR and FD-R: FD-DR excels with accurate nuisances and sufficient overlap; FD-R is more robust when overlap is weak or nuisances are noisy—consistent with the theoretical assessment in §4.1.

Highlights & Insights

• The "Transplant + Reparameterization" paradigm is clean: Instead of inventing entirely new estimators, the authors translate the front-door problem into solved back-door R-Learner sub-problems and use pseudo-outcomes/orthogonal losses to ensure debiasedness, reusing mature theoretical tools.
• pseudo-g is the crowning touch: It precisely identifies and eliminates a hidden source of slow convergence—where the combined term's error is bottlenecked by the propensity score—enabling FD-R to reach quasi-oracle rates.
• Complementary rather than competitive: §4.1 provides practitioner guidance (use FD-DR for strong overlap, FD-R for weak overlap) and proposes an adaptive routing mechanism based on overlap levels.
• Model-agnostic: Both nuisances and targets can use any off-the-shelf ML (XGBoost used in experiments), lowering the barrier for deployment.

Limitations & Future Work

• Reliance on positivity (overlap) assumptions: Variance inflates as \(e(X|C)\) or \(q(Z|XC)\) approach 0/1 (especially for FD-DR). The authors suggest overlap diagnostics, ratio stabilization, and overlap-aware uncertainty, planning an adaptive mechanism for automatic routing to FD-R.
• Limited to binary mediator \(Z\): The theory is established for binary \(Z\). Extending this to continuous or multi-dimensional mediators—common in real-world scenarios—is a clear future direction.
• Lack of direct comparison with deep heterogeneous front-door baselines: Experiments mainly compare against plug-in; a side-by-side comparison of debiasedness gains against methods like LobsterNet is missing.

Related Work & Insights

• Back-door Debiasing Lineage: AIPW (Robins 1994), TMLE (van der Laan), DR-Learner (Kennedy 2023), R-Learner (Nie & Wager 2021), orthogonal statistical learning (Foster & Syrgkanis 2023), DML (Chernozhukov 2018)—this work effectively migrates DR/R-Learners to the front-door setting.
• Front-Door Average Effect Debiasing: Fulcher 2019 (Doubly Robust FD-ATE), Guo 2023 (one-step/TMLE), Jung 2024 (Unified Covariate Adjustment); deep scalable but non-debiased approaches by Xu & Gretton 2022 / Xu 2024.
• Heterogeneous Front-Door: LobsterNet (multi-task neural net) by Chen 2025 is the most relevant prior work; this paper fills the gap by adding debiasedness.
• Insight: When an identification formula is complex, "reparameterizing into solved standard sub-problems and then designing orthogonal pseudo-outcomes" is a generalizable recipe for other identification criteria (e.g., napkin/general ID).

Rating

• Novelty: ⭐⭐⭐⭐ Simultaneously satisfies "front-door + heterogeneity + debiasedness" for the first time. The pseudo-g decoupling and reparameterization are substantial new constructions while following a mature toolset migration path.
• Experimental Thoroughness: ⭐⭐⭐ Synthetic experiments systematically validate the theory across four regimes. The FARS case study is convincing, but a direct comparison with deep heterogeneous front-door baselines is missing.
• Writing Quality: ⭐⭐⭐⭐ Clear theoretical derivations. The 3D positioning in Table 1c and the practical guidance in §4.1 are helpful and improve readability.
• Value: ⭐⭐⭐⭐ Provides model-agnostic, individualized causal estimation tools with fast convergence guarantees for observational scenarios (policy, healthcare, platforms) where latent confounding exists but compliant mediators are available.

Related Papers

• [ICLR 2026] ALM-MTA: Front-Door Causal Multi-Touch Attribution Method for Creator-Ecosystem Optimization
• [ICLR 2026] Learning Exposure Mapping Functions for Inferring Heterogeneous Peer Effects
• [ICLR 2026] GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes
• [ICLR 2026] Topological Causal Effects
• [ICLR 2026] Matching without Group Barrier for Heterogeneous Treatment Effect Estimation