The fundamental challenge of causal inference is unobserved confounding. When hidden variables influence both the treatment and the outcome, the observed data alone cannot pinpoint a single, unique causal effect. Traditional approaches bypass this by making strong, often untestable assumptions—such as assuming that all confounders have been measured.
However, a rapidly growing body of research takes a more honest approach: partial identification. Instead of forcing the model to output a single, potentially biased point estimate, partial identification seeks to find the exact mathematical range—the upper and lower bounds—of all causal effects that are perfectly consistent with the observed data.
The core insight uniting this field is that finding these bounds is fundamentally a constrained optimization problem. By treating the unknown causal query as an objective function to be maximized and minimized, and the observed data and structural assumptions as constraints, researchers can map out the precise boundaries of our causal knowledge.
The Geometry of Feasible Causal Worlds
To understand why optimization works for causal bounding, imagine the space of all possible Structural Causal Models (SCMs). Each specific SCM represents a distinct "causal world" with its own rules for how variables interact.
When we collect observational data, we rule out millions of these worlds. The remaining SCMs form a feasible region: a set of causal models that perfectly replicate the joint distribution we observe in reality.
Space of All Causal Models
┌───────────────────────────────┐
│ ┌───────────────────────┐ │
│ │ Feasible Region │ │
│ │ (Matches Data P(X,Y))│ │
│ │ │ │
│ │ ▲ Maximize Query │ │
│ │ │ ─── Upper Bound │ │
│ │ │ │ │
│ │ ▼ Minimize Query │ │
│ │ ─── Lower Bound │ │
│ └───────────────────────┘ │
└───────────────────────────────┘
To find the tightest possible range (known as sharp bounds) for a causal query—such as the Average Causal Effect (ACE) or a counterfactual probability—we simply search within this feasible region. The maximum value the query can take within this region becomes our upper bound; the minimum value becomes our lower bound. As shown by Kevin Muyuan Xia et al. in their work on Neural Causal Models for Counterfactual Identification and Estimation, if the maximum and minimum values converge to the exact same number, the query is uniquely identifiable, completely bypassing the need for symbolic algebraic derivations like do-calculus.
Discrete Spaces and Linear Programming
When dealing with discrete variables, this optimization landscape becomes remarkably well-behaved. The behavior of any individual in a discrete causal graph can be categorized into a finite number of "response types" (e.g., an individual who always gets sick regardless of treatment, or an individual who only gets sick if treated).
Because probabilities of these response types must sum to one and must map directly to the observed data, the constraints on our feasible region are entirely linear. Consequently, finding the sharpest causal bounds reduces to a standard Linear Programming (LP) problem.
This LP framework provides a rigorous mathematical toolkit across diverse scenarios:
- Truncation by Death: In clinical trials where outcomes cannot be observed if a patient dies early, Linbo Wang 0001 et al. demonstrated in Causal analysis of ordinal treatments and binary outcomes under truncation by death that LP can find the narrowest range of treatment effects across multiple dosage levels.
- Measurement Error: When data is imperfectly recorded, Noam Finkelstein et al. showed in Partial Identifiability in Discrete Data with Measurement Error that measurement error properties can be embedded directly as linear constraints, preserving the validity of the bounds.
- Continuous Outcomes: Even when outcomes are continuous, certain structural assumptions allow researchers to discretize or restrict the optimization space. For instance, Junzhe Zhang et al. developed tractable linear programs to bound causal effects for continuous outcomes under non-compliance in Bounding Causal Effects on Continuous Outcome.
Scaling Up: Continuous Variables and Neural Search
While LP is highly effective for low-dimensional discrete data, it suffers from the curse of dimensionality. If variables are continuous or high-dimensional, the number of response types becomes infinite, rendering standard LP solvers useless. To overcome this, researchers have introduced three major paradigms:
1. Polynomial Programming and Optimal Transport
Instead of linear equations, complex counterfactual queries can be framed as higher-order optimization problems. Jin Tian et al. proved in Partial Counterfactual Identification from Observational and Experimental Data that any counterfactual query in a discrete model can be represented using a finite set of unobserved states, transforming the bounding task into a polynomial program.
Alternatively, the problem can be viewed geometrically through the lens of fluid dynamics. Zijun Gao et al. introduced a framework in Bridging Multiple Worlds: Multi-marginal Optimal Transport for Causal Partial-identification Problem that treats the causal query as a cost function and the observed data as marginal constraints, utilizing Multi-marginal Optimal Transport (MOT) to calculate mathematically sharp intervals.
2. Generative Adversarial Networks (GANs)
To handle continuous variables without restrictive assumptions, the authors of A Generative Adversarial Framework for Bounding Confounded Causal Effects turned optimization into a game.
┌────────────────┐ ┌──────────────────┐
│ Generator │ │ Discriminator │
│ (Explores SCMs │────────>│ (Ensures Match │
│ to Min/Max) │ │ with Obs Data) │
└────────────────┘ └──────────────────┘
In this framework, a generator network acts as a universal function approximator that explores the space of possible SCMs, explicitly trying to maximize or minimize the target causal metric. Meanwhile, a discriminator ensures that the generated data matches the true observed joint distribution.
3. Regularized Neural Causal Models
Neural networks can also be optimized directly to find bounds. However, because neural networks are highly flexible, they run the risk of overfitting and hallucinating invalid causal worlds. To solve this, Jiyuan Tan et al. established the formal consistency of these methods in Consistency of Neural Causal Partial Identification, proving that enforcing strict Lipschitz regularization during training guarantees that the neural network converges to the true mathematical bounds as data increases.
Soft Constraints and Robustness
In practice, our knowledge of the system is rarely perfect. We might have noisy data, or our instrumental variables (IVs) might slightly violate the required causal assumptions. If we enforce hard constraints based on faulty assumptions, our optimization problem might return an empty feasible region or confidently incorrect bounds.
To address this, recent frameworks incorporate "soft" or distributionally robust constraints:
- Information-Theoretic Sensitivity: Ziwei Jiang et al. introduced Conditional Common Entropy (CCE) in Conditional Common Entropy for Instrumental Variable Testing and Partial Identification. By adding an entropy-based sensitivity parameter to the optimization framework, they allow the model to relax strict IV assumptions, narrowing the range of effects based on how much violation is tolerated.
- Minimax Robustness: When covariates are noisy, Wenshuo Guo et al. proposed a minimax optimization framework in Partial Identification with Noisy Covariates: A Robust Optimization Approach. This approach searches for the worst-case and best-case treatment effects across a defined set of distributions, ensuring the true causal effect remains captured even under measurement noise.
Putting Bounds to Work
Why settle for a range when decision-makers want a single answer? The power of partial identification is that a range is often more than enough to make optimal, safe choices.
Safe Reinforcement Learning and Bandits
In reinforcement learning, "causal confusion" occurs when an agent misinterprets a correlation as a causal link due to hidden environmental factors. By calculating causal bounds on unknown reward and transition functions, agents can avoid catastrophic actions.
For example, Elias Bareinboim et al. utilized partial identification in Causal Imitation for Markov Decision Processes: a Partial Identification Approach to achieve robust imitation learning in confounded environments. Similarly, in mobile health applications, Yi Shen et al. formulated risk estimation as a mixed-integer linear program in Risk-Averse Multi-Armed Bandits with Unobserved Confounders: A Case Study in Emotion Regulation in Mobile Health, computing tight mathematical guarantees on the Conditional Value at Risk (CVaR) to guide safe exploration.
Training Models on Bounds
Perhaps the most elegant application is using these mathematical bounds as labels to train downstream machine learning models. In the "unit selection problem"—where a company wants to target customers who will buy a product only if given an incentive—the true benefit function is unidentifiable.
As demonstrated by Ang Li et al. in Unit Selection: Learning Benefit Function from Finite Population Data, by calculating the mathematical upper and lower bounds of the benefit function via optimization, one can train a machine learning model directly on these intervals. The model learns to generalize the underlying causal structure across complex characteristic spaces, proving that we do not need point identification to make precise, causally-informed decisions.
Key Foundational Work
Counterfactual Probabilities: Computational Methods, Bounds and Applications (1994) — contributed 2 concept elements to this direction.
The Evaluation of Causal Effects in Studies with an Unobserved Exposure/Outcome Variable: Bounds and Identification (2012) — contributed 1 concept elements to this direction.