Proximal Identification and Estimation in Front-Door Causal Structures with Unobserved Confounding of the Mediator

Abstract: Unobserved confounding is a fundamental obstacle in causal inference problems. In the graphical modeling literature, a general theory has been developed that allows identification in the presence of hidden variables, with some limitations. In particular, Pearl's celebrated front-door criterion allows nonparametric identification in the presence of unobserved common causes of the treatment and the outcome, however it requires the presence of an unconfounded variable that mediates all causal influence from the treatment to the outcome. This stringent requirement limits the applicability of the front-door criterion. We propose proximal generalizations of the front-door criterion, allowing both arbitrary treatment/outcome confounding, and unobserved confounders of the mediator, provided informative proxies for the latter type of confounders are observed. In addition to deriving three new identification strategies in this setting, we provide plug-in and influence function-based estimation strategies for the resulting functionals, and evaluate their performance through simulations.

• The paper demonstrates three novel identification functionals for p(Y(a)) by extending proximal causal inference to handle unobserved mediator confounding.
• It develops both plug-in and influence function-based estimators with theoretical guarantees and multiple robustness, reducing bias relative to traditional front-door methods.
• Simulations in binary and mixed data settings validate the estimators, confirming consistency and efficiency even when standard assumptions are relaxed.

Proximal Identification and Estimation in Front-Door Causal Structures with Unobserved Confounding of the Mediator

Introduction and Motivation

This paper investigates a core challenge in graphical causal inference: the non-identifiability of causal effects in the presence of unmeasured confounders affecting not only the treatment ($A$) and outcome ($Y$), but also the mediator ($M$). Pearl's front-door criterion permits identification of $p(Y(a))$ under observed mediators and arbitrary confounding between $A$ and $Y$, yet critically requires the absence of mediator-outcome or mediator-treatment confounding. Empirically, this assumption is rarely satisfied, since mediators are often subject to their own unmeasured confounders.

The authors formally extend proximal causal inference methodology to the "composite bow graph" setting, which augments the conventional front-door graph to allow unmeasured confounders of the mediator provided suitable proxies are available. The central contribution is a rigorous derivation of three distinct identification functionals for $p(Y(a))$ leveraging observed proxy variables—each corresponding to different structural and independence assumptions. Estimation strategies, including plug-in and influence function-based estimators, are developed alongside theoretical guarantees. Simulations empirically support the theoretical advances.

Technical Context and Preliminaries

The setting is a nonparametric causal graphical model described by an acyclic directed mixed graph (ADMG) expressing the dependencies among treatment ($A$), mediator ($M$), outcome ($Y$), the unobserved confounder ($Y$0), and observed proxies for $Y$1, denoted $Y$2 and $Y$3. The unmeasured $Y$4 affects $Y$5, $Y$6, and $Y$7, producing nonidentifiability under classical identification algorithms (e.g., the ID algorithm [tianIdentificationCausalEffects2002], [shpitser06id]). In such models, even the presence of an observed mediator is insufficient to restore identifiability by the classic front-door formula.

Proximal identification leverages the existence of observable proxies ($Y$8) for $Y$9 that meet certain conditional independence and completeness constraints, enabling identification of functionals otherwise intractable due to the hidden confounding structure. The analysis builds upon prior work on nonparametric identification using proxies [miao2018identifying], [kuroki2014measurement], and extends recent developments on proximal algorithms for structural learning [shpitser2023proximal].

Identification Results

Assumption Set 1: Proximal Front-Door via Sequential Bridge Functions

The first identification strategy assumes a set of conditional independence conditions (Assumption Set 1) allowing for $M$0 to be associated with $M$1, and $M$2 to be associated with $M$3. Given completeness conditions on the conditional distributions $M$4 and solutions to a sequence of Fredholm integral equations, one can recursively construct bridge functions ($M$5, $M$6) expressing the distribution $M$7 and $M$8 as mixtures over their proxies.

The identified functional takes the form:

$M$9

where $p(Y(a))$0 is the solution to the aforementioned Fredholm equations.

Assumption Set 2: Alternative Proxy Configuration

A non-nested identification region is provided by Assumption Set 2, which assumes distinct conditional independence restrictions (notably, $p(Y(a))$1) suitable for distinct data-generating mechanisms. Under analogous completeness and bridge conditions (here, via functions $p(Y(a))$2, $p(Y(a))$3), identification is achieved by the functional:

$p(Y(a))$4

Both Assumption Set 1 and 2 functionals require the existence and computation of solutions to Fredholm integral equations, generalizing previous approaches in instrumental variable literature [newey03instrumental], and adapting them for this richer structural context.

Assumption Set 3: Full-Law Recovery via Kruskal-Type Arguments

The third identification regime, built on more restrictive structural assumptions (including mutual independence of $p(Y(a))$5 given $p(Y(a))$6 and strong completeness/distinctness on the support of $p(Y(a))$7), enables identification of the entire joint law up to label-swapping of $p(Y(a))$8. This allows not only $p(Y(a))$9 but all full-data functionals to be computed. The underlying identification employs tensor decomposition and spectral methods from latent variable modeling [kruskal1977three], [allman2015parameter], and, in practical terms, reduces to eigendecomposition tasks in finite-support settings. Importantly, this regime supports parametric and nonparametric estimation approaches, including the construction of efficient influence-function based estimators.

Estimation Procedures

For each identification strategy, the paper develops plug-in estimators, which—under correct model specification and standard regularity—are consistent for their respective functionals. For Assumption Set 3, a nonparametric influence function-based estimator is also provided, leveraging semiparametric theory to achieve $A$0-rate efficiency even when high-dimensional nuisance functions are estimated at slower rates (cf. the double/debiased machine learning literature [chernozhukov_2018], [KennedyTutorial]).

The estimator admits a form of multiple robustness: it remains consistent when certain combinations of the involved nuisance functions are misspecified, as long as some "sufficient" subset is consistently estimated (Theorem~\ref{thm:multiple_robustness}). This property is rigorously established and matches recent theoretical best practices for semiparametric inference under partial identification.

Simulation results substantiate the finite-sample performance of all developed estimators under both binary and mixed data-generating processes. Notably, functionals based on the (incorrect) front-door estimator retain substantial bias and variance even at moderate sample sizes, validating the necessity of the more elaborate proximal approaches.

Empirical Results

Numerical experiments were conducted for both finite-support (binary) and mixed-type (binary + continuous) data-generating processes simulating the structure of the composite bow graph with unobserved confounding. Plug-in and influence-function based estimators demonstrate substantial bias reduction and consistency relative to the mis-specified front-door estimator, with all proposed estimators converging to the ground-truth ACE as sample size increases. The influence function-based estimator retains consistency under certain forms of model misspecification, confirming both theoretical and practical claims.

Implications and Future Prospects

The principal implication of this work is the removal of a major barrier to the empirical application of front-door identification in causal inference: the assumption of an unconfounded mediator. By broadening the class of front-door-like structures to those where the primary restriction is on the availability and informativeness of observable proxies, the authors enable identification and estimation in settings previously regarded as intractably confounded.

Theoretically, this work rigorously connects advances in proximal causal identification with older tensor decomposition and measurement error results, yielding a comprehensive picture of the conditions required for identification in the presence of unobserved mediator confounding. Practically, the approach grants applied researchers new tools to leverage auxiliary variables for causal mediation analysis, provided the completeness and independence assumptions can be justified or approximately verified.

Future research directions include generalization to continuous $A$1 beyond current bridge solution constructs, development of robust methods for estimating the requisite bridge functions in high-dimensional or sparse data regimes, and exploration of nonparametric/regularization-based strategies for solving the required integral equations in more complex mediator-outcome spaces. The framework may also be extended to longitudinal mediation and dynamic treatment regimes.

Conclusion

This paper systematically advances the identification and estimation of causal effects in front-door-like structures compromised by unobserved mediator confounding. By leveraging informative proxies and formalizing requisite independence and completeness conditions, rigorous identification results are derived and accompanied by practical, robust estimators with demonstrable empirical performance. This development closes a critical gap in proximal causal inference and opens new avenues for empirical mediation analysis in the presence of latent variables.

References:

• "Identifying causal effects with proxy variables of an unmeasured confounder" [miao2018identifying]
• "Measurement bias and effect restoration in causal inference" [kuroki2014measurement]
• "Parameter Identifiability of Discrete Bayesian Networks with Hidden Variables" [allman2015parameter]
• "The proximal id algorithm" [shpitser2023proximal]
• "On the Identification of Causal Effects" [tianIdentificationCausalEffects2002]
• "Identification of Joint Interventional Distributions in Recursive Semi-Markovian Causal Models" [shpitser06id]
• "Three-way arrays: rank and uniqueness of trilinear decompositions, with applications to arithmetic complexity and statistics" [kruskal1977three]
• "Double/debiased machine learning for treatment and structural parameters" [chernozhukov_2018]