Optimization for Partial Identification of Causal Effects

This direction is split into other directions: constrained causal optimization. Its last version stays readable and citable here.

Version 1 of 173 papers97 conceptsfrozen 2026-10-04

In an ideal world, causal inference provides a single, precise number: "Taking this drug reduces the risk of heart disease by exactly 12%." However, in the real world, unobserved confounders—hidden factors that influence both the treatment and the outcome—often make such "point identification" impossible. When we cannot see every variable, the data is consistent with a wide range of possible causal effects.

Rather than making untestable assumptions to force a single answer, a growing body of research treats causal inference as an optimization problem. By framing the search for causal effects as a mathematical program, researchers can find the partial identification bounds: the tightest possible interval that is guaranteed to contain the true effect.

The Geometry of Uncertainty

The core intuition behind this approach is that while we don't know the exact causal mechanism, we do know that any valid mechanism must be consistent with the data we can observe.

In discrete settings, Jin Tian and Judea Pearl showed in Probabilities of Causation: Bounds and Identification that this consistency can be expressed as a set of linear constraints. They reimagined unobserved variables not as mysterious clouds of noise, but as "response functions"—deterministic mappings that dictate how an individual reacts to a treatment.

By treating the probability of these response functions as variables in a Linear Programming (LP) problem, the search for a causal effect becomes a search for the minimum and maximum values of an objective function within a "polytope" of feasible causal models. As Guilherme Duarte and colleagues explain in An Automated Approach to Causal Inference in Discrete Settings, this geometric strategy ensures that the resulting bounds are "sharp"—meaning they use every scrap of available information to narrow the range of uncertainty as much as mathematically possible.

Simplifying the Invisible

A major hurdle in this field is that unobserved variables can theoretically be infinitely complex. How can you run an optimization over something you can't even define?

Jin Tian, Junzhe Zhang, and Elias Bareinboim solved this by proving the existence of canonical partitions. In Partial Counterfactual Identification from Observational and Experimental Data, they demonstrate that any unobserved variable, no matter how complex, can be simplified into a finite, discrete state space without losing any counterfactual information.

This realization transforms a seemingly impossible infinite-dimensional problem into a solvable polynomial programming task. It allows researchers to combine disparate data sources—such as observational studies and randomized controlled trials—to shrink the bounds of counterfactual queries, such as "What is the probability that the treatment was both necessary and sufficient for the outcome?"

Scaling to the Continuous World

While linear and polynomial programming work well for categorical data (like "recovered" vs. "not recovered"), they struggle with continuous variables like drug dosages or blood pressure levels. To handle this, researchers have turned to the "universal function approximation" power of neural networks.

In A Generative Adversarial Framework for Bounding Confounded Causal Effects, Yaowei Hu and Xintao Wu introduce an adversarial approach. They use a Generator to explore the space of possible causal models and a Discriminator to ensure those models stay consistent with the observed data. By training the generator to maximize or minimize a causal metric, they can "push" the boundaries of the effect to find its limits in high-dimensional spaces.

However, neural networks can be "too flexible," sometimes hallucinating unrealistic, spiky response curves to satisfy constraints. Jiyuan Tan and Jose Blanchet argue in Consistency of Neural Causal Partial Identification that strict Lipschitz regularization is required. By limiting how fast the model's output can change, they ensure that the neural network converges to the true causal bounds as more data becomes available, rather than finding "cheating" solutions.

From Theory to Actionable Recourse

The shift from point estimates to optimization-based bounds has profound implications for high-stakes decision-making. In Algorithmic Recourse in Partially and Fully Confounded Settings Through Bounding Counterfactual Effects, Julius von Kügelgen and Bernhard Schölkopf apply these bounds to "algorithmic recourse"—the advice an AI gives you to change an outcome (e.g., "increase your income by $5,000 to get the loan").

Traditional recourse methods assume they know the exact causal model of the world. If they are wrong, their advice might fail. By using partial identification, these systems can instead provide "robust" recommendations: actions that are guaranteed to work across all possible causal models consistent with the data. This moves causal inference away from the fragile search for a single "truth" and toward a more honest, interval-based quantification of what we can and cannot know.

Papers in this version

Showing 60 of 73 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_optimization_for_partial_identification__01102d1d,
  title = {Optimization for Partial Identification of Causal Effects},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-for-partial-identification-of-causal-effects/rdr_01102d1d4c5c48b1a64af893dfe33594},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_optimization_for_partial_identification__01102d1d_v1,
  title = {Optimization for Partial Identification of Causal Effects},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-for-partial-identification-of-causal-effects/rdr_01102d1d4c5c48b1a64af893dfe33594/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}