Partial Identification Framework

• Partial identification is a statistical framework that defines parameter bounds when data and model constraints do not yield unique estimates.
• It employs methodologies like moment inequalities, linear programming, and support functions to derive sharp, nonparametric bounds on population quantities.
• The framework is applied in causal inference, missing data analysis, and weak supervision, offering robust sensitivity analyses and computational tractability.

Partial identification is a statistical and econometric framework that formalizes situations where the available data and assumptions do not uniquely determine a parameter of interest but instead restrict it to lie within a set, known as the identified set. Unlike classical point identification, which requires that the observable distribution uniquely pin down the parameter, partial identification recognizes the empirical and structural sources of non-identifiability—such as missing data, unmeasured confounding, support violations, or weak proxy information—and leverages all available constraints to deliver sharp, often nonparametric, bounds on population quantities. Modern developments employ tools from convex analysis, linear and convex programming, optimal transport, and random set theory to obtain, estimate, and make inference about these non-singleton intervals, with broad applications across causal inference, treatment effects, missing data, panel and network models, and program evaluation.

1. Formal Definitions and Foundational Principles

Partial identification replaces the classical mapping between a model-implied observable distribution and parameters—$M(\theta) = P$ iff $\theta = \theta_0$—with the weaker requirement that $M(\theta)$ contains $P$, yielding an identified set

$\Theta_I = \{\theta : P \in M(\theta)\}.$

This set may be a strict subset of the parameter space, possibly even a singleton under point identification, but generically forms a set-valued map in the absence of sufficient information. Under this framework, observationally equivalent parameterizations characterize regions rather than points, formalized via random closed sets $W(\omega)$ whose population-level law gives rise to $\Theta_I$ (Molinari, 2020).

Sharp (or minimal) bounds refer to the smallest interval containing the parameter consistent with observed data and maintained model assumptions. Any further tightening requires additional restrictions.

2. General Methodologies and Estimation Frameworks

Sharp bounds are often derived via systems of analytic inequalities, moment inequalities, or conditional constraints. Common principles include:

• Moment (In)equalities: For $\theta$ and moment maps $m_j$, $$\Theta_I = \{\theta : \mathbb{E}_P[m_j(X;\theta)] \leq 0,\, \forall j\}$$.
• Conditional Linear Programming (CLP): Bounds on parameters are the expectations of optimal values of LPs subject to data-driven linear constraints, possibly varying with covariates. This approach is central for problems such as MNAR estimation, IV bounds, and joint potential outcome distributions, where covariate stratification preserves sharpness (Ben-Michael, 13 Jun 2025).
• Random Set and Support Functions: Convex identified sets are characterized by their support functions $\theta = \theta_0$0, with inference reduced to functional estimation on the unit sphere.

Estimation proceeds by solving plug-in or sample versions of the analytic characterizations—plugging empirical means into the bounding functionals (e.g., Lee bounds or Fréchet bounds), or (for more complex problems) numerically optimizing over feasible sets defined by conditional constraints or transport couplings.

3. Application Scenarios and Key Model Classes

3.1. Missing Data and MNAR Mechanisms

With outcome $\theta = \theta_0$1 missing not at random ($\theta = \theta_0$2 observed), $\theta = \theta_0$3 is not identified without further restrictions. The classical bound is

$\theta = \theta_0$4

where $\theta = \theta_0$5 is the known support (Molinari, 2020). Recent advances incorporate weak shadow variables—predictions from pretrained models under the assumption $\theta = \theta_0$6—to yield much tighter bounds via LP formulations. When the matrix encoding observed joint structures with shadow variables is full-rank, point identification is achieved; otherwise, the feasible set of LP solutions provides set identification (Chen et al., 17 Feb 2026).

3.2. Partial-Label Learning and Weak Supervision

In PLL, only sets $\theta = \theta_0$7 containing the true label $\theta = \theta_0$8 are observed per example, not the true label. Partial-identification arises since only the minimum risk over $\theta = \theta_0$9 is identifiable for each observation. Progressive algorithms such as PRODEN assign soft label weights and iteratively update classifier and weights to minimize a surrogate risk estimator, converging to the Bayes rule when ambiguity is not maximal (Lv et al., 2020).

For weakly supervised performance evaluation, model metrics (accuracy, precision, recall) are bounded nonparametrically by classical Fréchet-Hoeffding inequalities using the observable marginals for predicted labels and weak labels, without requiring access to ground-truth (Polo et al., 2023).

3.3. Causal Inference with Unmeasured Confounding or Weak Proxies

Proxy and negative control methods relax the requirement that point-identification hinges on untestable completeness conditions. Proxy-based frameworks leverage weak shadow or bridge variables (possibly lacking completeness) to bound causal effects under ambiguous relationships between observed proxies and unobserved confounders. Bounds may be computed via min/max functionals or LPs, and are often non-differentiable but can be smoothed for bootstrap inference (Ghassami et al., 2023).

For multi-treatment, multi-outcome problems with latent factor confounding, the identified region for a vector of causal effects exploits the geometry of the possible bias terms, with external information on some estimands (such as effect-size restrictions or negative controls) further tightening or collapsing bounds to a point (Kang et al., 2023).

3.4. Panel Models and Support Violations

In nonlinear semiparametric panels with incomplete support for sufficient indices, point identification of average structural functions or average partial effects fails. However, combining integrals over observed support with best–worst-case placements of mass outside observed support yields sharp bounds (Liu et al., 2021).

3.5. Experimental and Policy Evaluation Settings

Under complex experiment designs (stratified randomization, attrition), standard point-identifying strategies are invalid. Generalizations of the Lee bounds, using tailored trimming and inverse probability weighting, provide valid sharp intervals for treatment effects (Ferman et al., 18 Jan 2026).

Policy learning with partial identification replaces the objective function by the lower bound (minimin value) over all feasible distributions, typically obtained by solving a CLP at each covariate value and integrating across the observed covariate distribution (Ben-Michael, 13 Jun 2025).

4. Computational Strategies and Inference

Most modern frameworks cast the identified set as the solution to linear or convex programs with constraints encoding all available structure—often conditional on covariates. Efficient algorithms include:

• LP solvers for finite and conditional programs: e.g., shrinking the feasible region with data-driven covariate or auxiliary variable constraints (Chen et al., 17 Feb 2026, Ben-Michael, 13 Jun 2025).
• Sieve or smooth convex optimization for transport/coupling problems: e.g., when combining two samples linked only by a proxy variable (Li, 12 Apr 2026).
• Entropy regularization for smooth approximation: resulting in differentiable bounds and scalable optimization (Ben-Michael, 13 Jun 2025).
• Orthogonal pseudo-outcome construction: for estimation of sharp bounds (even when bounds are functionals of unknown or high-dimensional quantities) with quasi-oracle properties, as in network partial identification under exposure mapping uncertainty (Schröder et al., 3 Feb 2026).

Asymptotic normality and valid confidence intervals for the set endpoints can be achieved under regularity and margin conditions, using influence-function based plug-in variance estimators, cross-fitting, and smooth approximations to non-smooth min/max functionals.

5. Theoretical Properties: Sharpness, Efficiency, and Robustness

Sharpness of bounds is guaranteed when all constraints on feasible models are imposed. Point identification occurs in degenerate cases (e.g., when auxiliary or shadow variables provide invertible information), but generically, identified sets remain interval-valued. Set estimators are consistent in Hausdorff distance to the population set under mild regularity, and confidence sets on the projection (support function) admit uniform coverage (Molinari, 2020).

Robustness of conclusions is assessed via sensitivity parameters (e.g., allowed degree of unmeasured confounding or exposure misspecification), with breakdown frontier analysis mapping how much assumption violation a given qualitative inference (such as sign of effect) can withstand (Lanners et al., 30 May 2025).

6. Advances and Empirical Illustrations

Recent work demonstrates the practical gain from partial identification:

• Incorporating LLM-derived weak signals in MNAR missing data yields up to 83% reduction in interval width while retaining valid coverage under realistic missingness (Chen et al., 17 Feb 2026).
• In network causal inference under misspecified exposure mapping, sharp closed-form bounds can differentiate effect sign even under substantial ambiguity, with doubly robust estimators achieving quasi-oracle rates (Schröder et al., 3 Feb 2026).
• Combining programmatic weak signals in model validation settings, Fréchet bounds supply informative accuracy/precision bounds in high-dimensional industrial weak supervision pipelines (Polo et al., 2023).
• In multi-source causal fusion, interpretable sensitivity parameters delineate the robustness of findings across regions of assumption violation (Lanners et al., 30 May 2025).

7. Conceptual and Practical Implications

The partial identification framework systematizes and unifies the analysis of models with fundamentally non-unique solutions, focusing inference on what is empirically and structurally warranted. The approach replaces default reliance on untestable assumptions with transparent, data-driven uncertainty quantification, and offers computationally tractable methods for estimation, inference, and policy optimization in highly complex, high-dimensional, or weakly identified environments.

The framework continues to evolve toward flexible incorporation of modern machine learning (e.g., ML-generated proxies), expansion of finite-sample guarantee regimes (e.g., set-expansion estimators under ill-conditioning), and integration of stochastic or nonparametric modeling approaches (e.g., random set theory, semiparametric and Bayesian nonparametrics). Its reach now spans missing data, causal analysis with unmeasured confounding, partial-label learning, model evaluation without ground truth, complex survey and experimental designs, and high-dimensional decision-making.