Open Challenges in Statistical Decision Theory

• Statistical decision theory is a formal framework for making optimal choices under uncertainty by balancing sampling variability and model ambiguity.
• Key challenges include addressing model misspecification, handling partial identification, and ensuring robustness through adaptive and minimax-regret approaches.
• Emerging research targets scalable algorithms and integration with modern ML to enhance robust decision-making in complex, high-dimensional data environments.

Statistical decision theory provides a rigorous mathematical formalization for choosing actions under uncertainty, balancing sampling variability and model uncertainty. It underpins optimal experimental design, estimation, robust prediction, and decision-making for policy and science. Contemporary advances and complex data environments have foregrounded a suite of open challenges, ranging from foundational questions about model misspecification and identification, through high-dimensionality and robustness, to fundamental barriers in inference under ambiguity and interaction. The following sections synthesize major open problems and research frontiers identified in the recent literature.

1. Model Misspecification, Ambiguity, and the True State-Space

Statistical decision theory traditionally assumes that the true data-generating process lies within a specified model space. In practice, models are only approximations, and the true state may fall outside the class considered in estimation and inference. This gap—formalized as Θₘ ≠ Θ, with Θₘ the model space and Θ the state space of all plausible distributions—yields critical vulnerabilities in decision procedures designed under misspecification. Rules tuned to Θₘ can have arbitrarily large regret when evaluated over Θ, especially if nonresponse, unmeasured confounding, or population shift are present (Manski, 2019, Dominitz et al., 2024). A central open direction is to formalize, quantify, and control risk under such model misspecification, including development of decision functions robust to both sampling and model uncertainty.

Decision-making under ambiguity—a persistent rather than reducible uncertainty—necessitates robust (minimax or minimax-regret) rules, and frameworks that explicitly accommodate estimation uncertainty, ambiguity sets (e.g., likelihood neighborhoods in parameter space), and adversarial worst-case scenarios in dynamic or high-dimensional systems (Blesch et al., 2021). There remain open computational and conceptual questions about optimal criteria—Bayes, minimax, and minimax-regret—and their appropriateness in practical settings.

2. Identification, Partial Identification, and Ambiguity

Identification analysis (determining how tightly parameters can be estimated from observed data) imposes upper bounds on achievable decision performance. Partial identification, where only sets of feasible parameter values can be inferred, introduces ambiguity that interacts nontrivially with statistical uncertainty. Even when the sampling distribution is well characterized, the identified set Θ_I may be large or nonconvex, precluding pointwise optimality.

A pivotal open challenge is the development of decision rules—often randomized—that minimize maximal risk or regret over ambiguous identified sets, especially when these sets are described by complicated constraint systems or infinite-dimensional function sets (e.g., moment inequalities, shape restrictions) (Manski, 2022, Qiu et al., 25 Jan 2026). Existing theory provides closed-form minimax-regret solutions only in low-dimensional or highly symmetric settings. Open questions include:

• Structural characterization and existence/uniqueness of optimal rules for general partial identification models.
• Extension of minimax/minimax-regret methods to continuous action spaces, possibly infinite-dimensional.
• Efficient computation of decision rules, especially in high dimensions or under complex constraints.

Qiu and Stoye (Qiu et al., 25 Jan 2026) detail three broad areas for further research: theoretical generalization to richer identification complexities, scalable algorithms for large or infinite constraint systems, and practical, tractable bounds for policy designs with partial identification.

3. Robust Decision Theory: Contamination, Outliers, and Complex Data

Contemporary datasets often exhibit contamination, adversarial outliers, and distributional deviations. Huber's $\epsilon$-contamination model explicitly models data as a mixture of 'clean' and unknown 'contaminated' observations. The main open problem here is to construct estimators and decision procedures that adaptively attain minimax rates (in risk or error exponent) for a broad class of loss functions—not only those equivalent to total variation distance (Chen et al., 2015). For many metrics, such as the supremum norm in nonparametric estimation, extending adaptive procedures to unknown $\epsilon$ while preserving optimality remains unresolved.

Key open challenges include:

• Extending adaptive construction to loss functions not intrinsic to total variation, such as $\ell_\infty$, Wasserstein, or other divergences.
• Generalizing robust minimax estimation and testing to high-dimensional and nonparametric regimes, marrying robust optimization theory with learning-theoretic complexity.
• Efficient computational realization: algorithms must handle large covering numbers, model selection, and tuning without knowledge of contamination proportion or structure.

4. Computation, Algorithmic Scalability, and High-Dimensional Regimes

Statistical decision theory for modern data often entails action spaces, covariate structures, and model spaces that are high or infinite-dimensional. Direct optimization of risk or regret over these spaces is generally intractable. Notably, comprehensive out-of-sample (OOS) evaluation—risk assessment that averages over all possible samples and relevant populations—quickly becomes computationally prohibitive (Dominitz et al., 2024). Monte Carlo, grid-search, and surrogate optimization partially mitigate these issues, but scalable theoretical and algorithmic tools remain primitive.

Some core computational challenges:

• Developing scalable approximation algorithms (randomization, stochastic optimization, convex relaxations) for minimax and minimax-regret risk functions in large hypothesis spaces (Manski, 2022, Foster et al., 2021).
• Ensuring convergence and quality guarantees (e.g., $ε$-approximation) for randomized and saddle-point-based solvers in high or infinite dimensions (Qiu et al., 25 Jan 2026).
• Automated software frameworks for SDT-based evaluation and deployment in common prediction and policy-learning pipelines, especially for modern ML models (deep nets, ensembles).
• Untangling the relationships between statistical and decision-theoretic complexity measures, such as Rademacher complexity or the Decision-Estimation Coefficient (DEC) (Foster et al., 2021).

5. Statistical Decision Theory for Dynamic, Interactive, and Counterfactual Environments

Classical decision theory predominantly treats static action-selection. Dynamic, sequential, or interactive decision problems (multi-stage policies, reinforcement learning) introduce additional layers of complexity: adaptive randomness and partial observability.

Core open directions in this area are:

• Complexity theory for interactive statistical decision making, including tight characterizations of minimax regret in sequential and interactive settings. The Decision-Estimation Coefficient (DEC) provides necessary and sufficient conditions for sample-efficient interactive learning, but extending this framework to rich model classes (e.g., deep nets, general POMDPs), decentralized/multi-agent settings, and constrained batch/partial monitoring remains active research (Foster et al., 2021).
• Robust and ambiguity-aware dynamic programming: scalable robust Markov decision process (RMDP) algorithms that handle ambiguity sets beyond KL-balls, relax rectangularity, and address state-dependent data scarcity (Blesch et al., 2021).
• Counterfactual risk and loss: when evaluating and designing decision rules that account for hypothetical alternative actions and outcomes (“counterfactual loss”), the identification of risk is inextricably tied to additivity of the loss function and the feasibility of observing potential outcomes. Extending identification, estimation, and risk minimization to continuous actions, dynamic settings, and high-dimensional covariates remains an unsolved problem (Koch et al., 13 May 2025).

6. Statistical Decision Theory Under Model Selection, Hyperparameter Tuning, and Modern ML

The increasing complexity of ML models—deep nets, random forests, ensembles—and their validation protocols create new gaps between conventional OOS/CV-based model assessment and decision-theoretically meaningful guarantees. Key open areas:

• Formalizing the out-of-sample performance of ML algorithms, integrating over all sampling draws and state ambiguity, as opposed to ex post accuracy on held-out data or cross-validation. Minimizing decision-relevant regret as the core criterion for hyperparameter tuning and meta-model selection (Dominitz et al., 2024).
• Bridging local model-neighborhood robustness (e.g., Wasserstein balls, $f$-divergence sets) with global SDT formulations, particularly in the presence of high-dimensional covariate and state spaces.
• Empirical studies and theoretical characterizations of compositional sparsity and its effect on the statistical complexity of real-world ML problems.

7. Publication Bias, Sequential Learning, and Reproducibility

Sequential updating of beliefs, design of evidence thresholds (e.g., level-of-significance α in NP testing), and accumulation of evidence across studies are fundamentally complicated by publication and reporting biases. Open problems include:

• Design of decision protocols and reporting standards that support consistent Bayesian updating and guard against "researcher degrees of freedom" and selective publication (Pena, 2019).
• Integration of likelihood-ratio-based summary statistics and evidence-threshold sample-size planning into experimental design and reporting, replacing reliance on default significance thresholds.
• Aggregation and meta-analysis when studies differ in design, thresholds, or action sets, preserving decision-theoretic alignment and reproducibility.

Table: Principal Open Challenges in Statistical Decision Theory

Challenge Area	Representative Issue	Reference arXiv ID
Model ambiguity	Robustness under misspecification/ambiguity	(Manski, 2019, Blesch et al., 2021)
Partial identification	Minimax-regret/randomized rules for complex Θ_I	(Manski, 2022, Qiu et al., 25 Jan 2026)
Robust estimation/testing	Adaptive rates for metrics ≠ total variation	(Chen et al., 2015)
Computation & scalability	Efficient algorithms for SDT in high-dim/complex S	(Dominitz et al., 2024, Foster et al., 2021)
Dynamic & interactive SDT	Regret complexity in RL and sequential decisions	(Foster et al., 2021, Blesch et al., 2021)
Modern ML & hyperparameter	SDT-based evaluation/tuning of complex ML models	(Dominitz et al., 2024)
Sequential learning/bias	Meta-analysis, publication bias, reporting design	(Pena, 2019)
Counterfactual risk	Identification/estimation for multi-arm, time, non-binary, high-dim settings	(Koch et al., 13 May 2025)

The field continues to advance rapidly at the intersection of modern causal inference, robust statistics, computational optimization, and machine learning. Progress in these open areas will be prerequisite for ensuring that statistical decision theory delivers actionable, reliable, and computationally tractable foundations for research and policy in the presence of inherent and often irreducible uncertainty.