Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects

arXiv:2310.08115 · econ.EM, math.ST, stat.ME, stat.ML, stat.TH · Submitted 2026-08-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects".

Jane: The paper was written by Wenlong Ji, Lihua Lei and Asher Spector from Department of Statistics, Stanford University and Graduate School of Business and Department of Statistics, Stanford University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Summary: Tom: So, we've seen the title and the authors; now let's talk about what Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects actually does in a nutshell. The paper summarizes three specific examples of partially identified parameters, like Frechet-Hoeffding bounds or the variance of treatment effect heterogeneity.

Jane: These examples really show how this concept applies to different measures, from simple CDF thresholds to more complex statistics like the variance. It's not just one single number but a set defined by inequalities.

Lu: The key insight here is that all these estimands, regardless of their shape, fit into a single model structure that makes them tractable for this dual bound approach.

Meng: My concern is how this relates to real-world data collection; if we're using the framework for Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects, it’s because the traditional methods struggle with these kinds partially defined goals.

Lalam: The summary highlights that by transforming these complex estimands into a general mathematical form, we can apply this unified approach to estimate sharp bounds.

Improvements: Tom: The paper really shines in how it offers improvements over existing methods, particularly regarding model selection and robustness. It claims the method is "model-agnostic," which is a massive relief for researchers.

Jane: That means you don't need to assume your model of the conditional distribution of Y X, W is perfectly accurate; even if it might be misspecified, the bounds remain valid.

Lu: The improvement over previous work is that we can apply the multiplier bootstrap to select covariates and models without compromising that validity.

Meng: From an engineering standpoint, having a set of candidate models and choosing the best one—the tightest bound—is much more manageable than trying to force a single, rigid parametric model onto complex data.

Lalam: The ability Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects offers allows us to select the best outcome model while maintaining mathematical rigor is a massive step toward making robust AI applications in causal inference possible.

Conclusion: Tom: We've covered the core theory, but let's wrap up by summarizing what this method brings to the real world. The paper demonstrated that Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects is robust, computationally efficient, and highly flexible.

Jane: It’s a powerful tool because even though the results can be conservative or anti-conservative under misspecification, those that were accurate yield significantly sharper confidence intervals than existing methods.

Lu: The fact that this framework handles complex estimands—not just simple averages—is really what makes it so versatile for applying this to various economic and scientific datasets.

Meng: It’s a practical win because of the computational efficiency, allowing us to run these complex dual bounds analyses without requiring massive hardware or prohibitive amounts of time.

Lalam: For Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects, I believe the greatest impact is that it provides confidence in results even where certainty isn' a guarantee.

Tom: That’s a perfect way to summarize it all up. We have to thank the authors for this incredible work and hope you enjoyed our discussion of Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects.

Lu: I can't wait to see how this translates into real life applications in my research.

Meng: I’m excited to see how many production systems can now use these robust bounds in their decision making processes.

Lalam: This is a truly unifying concept that brings together theory and practical utility for the future of all disciplines.

Conclusion: Tom: So, we're wrapping up our discussion on "Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects," and I think it’s clear this paper has a huge amount to say about how we approach uncertainty in causal inference.

Jane: It’s not just about having a confidence interval, Tom; the paper is showing us *how* to build that interval reliably, even when the underlying data is messy or misspecified.

Lu: I see this as a major theoretical shift where practical constraints like model accuracy are simply no longer the primary bottleneck for achieving robust inference in high-dimensional systems.

Meng: From an engineering standpoint, I’m really impressed by how efficient this is—it' doesn't require us to run massive simulations just to get a viable dual bound estimate.

Lalam: It offers a deeply unifying approach that allows us to quantify the reliability of causal claims across different fields in a way that improves our overall societal confidence in data-driven decisions.

Tom: That’s right, Lalam, it makes the paper applicable almost everywhere by looking at the core structure of partial identification rather than some specific assumptions about your conditional distribution.

Jane: And Lu's point is that we' can move away from complex parametric assumptions and Meng's point is that we' can actually implement this in a real-time system today.

Lu: It’s certainly going to inspire a lot of creative solutions for modeling things where the true underlying mechanism isn’t known perfectly.

Meng: I think it just makes my job easier because we can use this framework as a baseline, even if we've only trained a basic AI model on the conditional distribution.

Lalam: We should all celebrate this advancement in "Model-Agnostic Covariate-Assisted Inference on Partially Identified Causal Effects" and look forward to how it helps us navigate complex data landscapes.

Tom: Absolutely, Lalam; we've got some really interesting papers lined up for next time that I think you guys will find fascinating too.

Wenlong Ji, Lihua Lei, Asher Spector

Department of Statistics, Stanford University · Graduate School of Business and Department of Statistics, Stanford University

econ.EM, math.ST, stat.ME, stat.ML, stat.TH

Submitted: 2026-08-24

Updated: 2026-08-25

Comments: 103 pages, 3 figures

Code: https://github.com/amspector100/dual_bounds_paper

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 73/100

The gist: Many causal parameters are "only partially identifiable" (Manski, 2003; Tamer, 2010).

Key concepts

Model-Agnostic
This means the inference method does not require perfect assumptions about the conditional distribution of Y given X and W. Even if the underlying model is misspecified, the resulting confidence bounds remain mathematically valid and robust for researchers.
Partially Identified Causal Effects
These are parameters that cannot be determined by a single number. Instead, they are defined as a set of values bounded by inequalities. Examples include variance estimates or CDF thresholds, making them complex to measure.
Covariate-Assisted Inference
This statistical approach improves inference by incorporating additional variables (covariates) into the analysis. It is used here to estimate sharp bounds and increase the reliability of causal claims in complex data sets.

Terminology

Summary

Many causal parameters are only partially identifiable (Manski, 2003; Tamer, 2010). Since we observe at most one outcome per subject in a randomized experiment, the joint law of potential outcomes is unidentifiable. While incorporating information from covariates X i in R p can substantially reduce the width of the partially identified set, inference remains challenging because the bounds typically depend delicately on the relationship between the outcome and the covariates.

The paper addresses this challenge by asking: "can we convert a working estimate of the conditional law of YX, W into inferential bounds on the sharp identified set [theta L, theta U] which are (i) sharp when the working estimate is consistent and (ii) conservative but valid when the working estimate is arbitrarily inaccurate?"

The authors propose a unified and model-agnostic inferential approach for a wide class of partially identified estimands. The core of this method is leveraging duality theory for optimal transport problems to convert any estimate P̂Y X,W of the conditional law into robust partial identification bounds and.

The method outputs:

  • Estimates L, U of the sharp bounds theta L, theta U.

  • Lower and upper confidence bounds,.

The method possesses four key properties:

  1. Uniform Validity: In randomized experiments, the approach can wrap around any estimates of the conditional distributions and provide uniformly valid inference, even if the initial estimates are arbitrarily inaccurate. This makes it model-agnostic. Formally, L and U are always conservatively biased: E[L] at most theta L and E[U] at least theta U.

  2. Tightness: If nuisance parameters are estimated at semiparametric rates, our estimator is asymptotically unbiased for the sharp partial identification bound.

  3. Model Selection: The method allows the analyst to use the multiplier bootstrap to select covariates and models without sacrificing validity, even if the true model is not selected.

  4. Computational Efficiency: The method is designed to be computationally efficient, even when X is high-dimensional and Y is continuous.

The paper establishes several formal results:

  • Theorem 3.1 (Uniform Validity): P(LCB at most theta L) at least 1 - alpha as n to infinity for P in PB.

  • Theorem 3.2 (Tightness): If strong duality holds, the first stage bias is bounded by the product of the errors in estimating (P*)(X) and nu*. This implies that if strong duality holds, sqrt n (L - theta L) = o(1).

  • Theorem 3.5 (Cross-fitting): When using cross-fitting, the resulting lower confidence bound LCB is valid under two conditions: either the dual variables are asymptotically deterministic, or the outcome model is sufficiently misspecified such that the first-stage bias is larger than n-1/2.

The method requires estimating conditional distributions P̂Y(0)X and P̂Y(1)X. While various methods exist, the authors recommend:

  • Cross-fitting: A technique used to recover the factor of two lost by sample splitting, leading to crossfit LCB = (L + swap L)/2.

  • Model Selection: The authors propose using the Gaussian multiplier bootstrap (Chernozhukov et al., 2013a) to select the tightest bound among K candidate estimates(k).

The method was tested in three empirical applications:

  1. Persuasion effects of political news (Gerber et al., 2009): The covariate-assisted dual bounds are more than twice as narrow as the covariate-free bounds, and they are more reliable than the covariate-assisted plug-in bounds.

  2. Estimating intensive margins (Carranza et al., 2022): For both logged and non-logged outcome measures, the covariate-assisted bounds are only about 60% as wide as the covariate-free bounds.

  3. 401k eligibility (Chernozhukov et al., 2018a): The dual lower bounds were found to be smaller than the ATE estimates, and they provide rigorous uncertainty quantification without assuming that the outcome model is accurate.

The paper concludes by noting that the method allows for extensions beyond causal inference (e.g, to linear programs) and discusses alternative computational strategies, such as using deep learning to learn the dual variables directly via Deep Dual Bounds.

Improvements for AI systems

As a diligent researcher, I must ensure that any implementation of this methodology yields statistically rigorous results. The core contribution of this paper is the decoupling of predictive accuracy from inferential validity—a critical distinction that standard machine learning often fails to make.

Below are the specific architectural and algorithmic improvements I propose for integrating these Dual Bounds into advanced AI systems, detailing exactly what each improvement enables.

The Improvement: Implement a post-processing layer that accepts any machine learning model's predicted conditional distribution (YX,W) and converts it into a mathematically rigorous set of confidence bounds [L, U] using the Kantorovich duality framework (Section 2.1).

What the System Can Do:

  • Guaranteed Validity: The system will provide Uniform Validity. Unlike standard ML methods, it guarantees that LCB at most theta L and UCB at least theta U, regardless of how inaccurate or poorly specified the underlying prediction model is.

  • Robustness: It moves beyond the plug-in fallacy, allowing the system to perform reliable inference even on observations where the conditional distribution is completely misspecified (e heteroskedasticity, etc.).

The Improvement: Integrate a mechanism to execute both Cross-Fitting and the Gaussian Multiplier Bootstrap (Section 2.3). This allows the system to test K candidate models ((k)) on the first fold (D 1) and select their corresponding dual bounds (theta L(k)).

What the System Can Do:

  • Optimal Model Selection: The system automatically selects a subset of covariates and/or a specific model architecture that yields the tightest possible lower bound k in [K] theta L(k). This eliminates manual tuning and ensures the most efficient use of information.

  • Adaptive Inference: It allows for Multi-Model Ensemble inference, combining results from multiple candidate models via the Multiplier Bootstrap to achieve a statistically sound aggregation of bounds.

The Improvement: Replace traditional infinite-dimensional optimization with a discretized Linear Programming (LP) approximation (Section 4.2). This allows the system to handle high-dimensional covariate spaces (X) without computational collapse.

What the System Can Do:

  • Scalability: The system can compute valid dual bounds for complex, continuous, or high-dimensional input features (X in R p) that would otherwise render traditional constrained optimization intractable.

The Improvement: Implement the Augmented Inverse Probability Weighting (AIPW) estimator combined with the Cross-Fitting technique (Section 3.4).

What the System Can Do:

  • Inference in Real-World Settings: The system provides a robust, asymptotically valid lower confidence bound (LCB) for Average Treatment Effects (ATE) even when propensity scores (pi(X)) are unknown and must be estimated on a separate data fold (D 1). This overcomes the severe limitations of existing methods in observational studies.

An AI system built upon these improvements will not merely predict outcomes; it will provide provably valid, model-agnostic uncertainty quantification around causal effect estimates. It transforms predictive models into a rigorous framework for decision-making by identifying the sharpest possible bounds on parameters, even when those models are flawed.

Sources

Related papers