The Identifiability Stability Gap in Causal Inference

37 concepts25 papers1 proposal

In causal inference, a celebrated milestone for any problem is proving identifiability. Identifiability means that if we had infinite data and perfect, noise-free probability distributions, we could write down a symbolic formula to uniquely recover a hidden causal effect.

But there is a massive gap between symbolic exactness and empirical reality: the identifiability-stability gap. When we feed real-world, finite, or noisy data into these mathematically sound formulas, the computations can break down spectacularly. A model can be perfectly identifiable in theory, yet completely unusable in practice because small observational errors explode into wildly inaccurate causal estimates.

Understanding why this happens—and how to fix it—requires looking under the hood of two major causal frameworks: proximal/instrumental variable methods and structural equation models.

The Operator Inversion Nightmare

When we cannot measure all confounders directly, we often rely on proxy variables or instrumental variables to chart a path around these "hidden ghosts" in our data. In proximal causal inference, we use treatment proxies ($Z$) and outcome proxies ($W$) to discover a hidden "bridge function" that accounts for unmeasured confounding.

Mathematically, finding this bridge function requires solving a Fredholm integral equation of the first kind. This structure appears across a wide range of causal problems, including nonparametric instrumental variable (NPIV) regression and off-policy evaluation in confounded environments, as discussed by Rui Miao et al. in Off-Policy Evaluation for Episodic Partially Observable Markov Decision Processes under Non-Parametric Models.

To understand why these equations are inherently ill-posed inverse problems, consider the intuition of image deblurring:

[True Causal Function] ---> (Conditional Expectation / Blur) ---> [Observed Data Distribution]
         h(X)                                                              P(Y|Z)

[True Causal Function] <--- (Operator Inversion / Deblur) <--- [Noisy Empirical Data]
       Unstable!                                                       P̂(Y|Z)

Taking a conditional expectation is a smoothing operator. It takes a sharp, underlying causal function $h(X)$ and averages it out to produce the observed data distribution $P(Y|Z)$. Solving the integral equation means running this process in reverse: we must "deblur" our observed data to reconstruct the true causal function.

Inverting a smoothing operator requires amplifying high-frequency details. If our empirical data contains even a tiny amount of statistical noise or measurement error, the inversion process amplifies that noise exponentially. As AmirEmad Ghassami et al. demonstrate in Minimax Kernel Machine Learning for a Class of Doubly Robust Functionals with Application to Proximal Causal Inference, this instability becomes severe when proxies are weakly correlated with hidden confounders. In these weak-proxy regimes, the underlying equations become so ill-posed that the variance of the causal estimate degrades completely.

This creates a profound theoretical challenge. As Yuri Fonseca et al. point out in Nonparametric Instrumental Variable Regression through Stochastic Approximate Gradients, an algorithm's optimization risk might converge beautifully to zero, yet the estimated function itself can remain wildly far from the true causal function. Traditional statistical guarantees of consistency often fail to protect against this hidden instability, a vulnerability shared by mediation analysis (Oliver Dukes et al. in Proximal mediation analysis), semiparametric estimators (Hongming Pu et al. in Semiparametric Proximal Causal Inference), and continuous instrumental variable settings (Yuta Kawakami et al. in Identification and Estimation of Conditional Average Partial Causal Effects via Instrumental Variable).

Condition Numbers and the Precision Trap

The same instability haunts graph-based models and Linear Structural Equation Models (LSEMs), even when the math reduces to matrix algebra rather than infinite-dimensional integrals.

Imagine trying to find the intersection point of two lines on a graph. If the lines cross at a sharp $90^\circ$ angle, a tiny nudge to either line barely moves the intersection point. But if the lines are nearly parallel, a microscopic nudge can slide the intersection point miles away.

In numerical linear algebra, this sensitivity to input perturbations is quantified by the condition number. When causal identification formulas require inverting near-singular probability matrices or evaluating complex algebraic terms (like square roots of empirical covariances), the condition number skyrockets.

Low Condition Number (Stable)          High Condition Number (Unstable)
             \   /                                   ___________
              \ /                               ____---======---____
               X                                --------------------
       (Precise Intersection)                  (Ambiguous Intersection)

In Stability of Causal Inference, Leonard J. Schulman et al. proved that standard causal identification algorithms can suffer from exponential noise amplification. As the number of variables in a causal graph grows, the precision required for your input data can quickly exceed physical limits.

Crucially, Spencer Gordon et al. showed in Condition number bounds for causal inference that multiple causal formulas can be algebraically identical under perfect data, yet exhibit completely different stability profiles when faced with noise. Choosing the wrong mathematically equivalent formula can mean the difference between a robust estimate and total computational failure.

To address this, researchers have begun moving beyond binary notions of identifiability. Karthik Abinav Sankararaman formalized this paradigm shift in Robust Identifiability in Linear Structural Equation Models of Causal Inference, arguing that we must demand robust identifiability—ensuring that our theoretical solutions are explicitly bounded against data perturbations.

Designing for Stability

To bridge the identifiability-stability gap, modern causal frameworks are replacing fragile, exact symbolic inversions with architectures designed to absorb and mitigate noise.

1. Regularization and Alternative Objectives

Instead of attempting direct operator inversion, newer methods restructure the optimization problem. Zihao Li et al. introduce Tikhonov regularization to instrumental variable regression in Regularized DeepIV with Model Selection, providing a stable alternative to the "instability nightmare" of unregularized minimax optimization. Similarly, Krikamol Muandet et al. bypass unstable two-stage regressions in Dual Instrumental Variable Regression by reformulating the problem as a smooth, convex-concave saddle-point optimization.

2. Adaptive Deep Feature Learning

Rather than relying on rigid, hand-crafted kernels that struggle with ill-conditioned data, deep learning approaches can adaptively smooth out the ill-posedness. Juno Kim et al. demonstrate in Optimality and Adaptivity of Deep Neural Features for Instrumental Variable Regression that deep neural networks excel at learning representations that adapt to "spiky" functions, achieving optimal convergence rates where traditional methods falter. This flexibility is leveraged by frameworks like Neural Maximum Moment Restriction (NMMR), introduced by Benjamin Kompa et al. in Deep Learning Methods for Proximal Inference via Maximum Moment Restriction, which uses neural networks to scale proximal inference to high-dimensional data like images without collapsing under numerical noise.

3. Closed-Form Structural Reformulations

Sometimes, the math itself can be re-engineered to minimize condition numbers. Liyuan Xu et al. show in Kernel Single Proxy Control for Deterministic Confounding that deriving specific closed-form kernel solutions can directly prevent the high conditioning numbers that typically cause estimation failures. By constraining the algebraic space of the solution, they ensure the estimation remains stable even when processing complex inputs.

By shifting the focus from what is theoretically identifiable to why certain formulations fail numerically, this line of research ensures that causal inference can be reliably deployed in messy, real-world environments.

Key Foundational Work

Uniform confidence bands for functions estimated nonparametrically with instrumental variables (2011) — contributed 1 concept elements to this direction.
Nonparametric Instrumental Variables Estimation of a Quantile Regression Model (2007) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 23 papers
  2. Version 1 · 2026-10-04 · 35 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_ill_posed_inverse_problems_numerical_ins_5b056a80,
  title = {ill-posed inverse problems (numerical instability in causal inference)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ill-posed-inverse-problems-numerical-instability-in-causal-inference/rdr_5b056a800ce64f648265669ef4e512a2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_ill_posed_inverse_problems_numerical_ins_5b056a80_v2,
  title = {ill-posed inverse problems (numerical instability in causal inference)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ill-posed-inverse-problems-numerical-instability-in-causal-inference/rdr_5b056a800ce64f648265669ef4e512a2/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake