Computational Identifiability

2026-06-19 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Researchers from New York University introduce "Computational Identifiability," a novel framework that redefines identifiability as a finite computational search for an empirical estimator, contrasting it with traditional theoretical identifiability that assumes idealized conditions like infinite data. This approach considers a target query identifiable if an estimator can be found empirically within a specified error tolerance, conditional on a prior distribution over parameters and the search procedure. The framework leverages a meta-prior over structural causal models (SCMs) and a hypothesis space of estimators, operating with finite sample sizes, a desired error tolerance ε, and a confidence bound δ. Experiments demonstrate its utility in practical scenarios, including identification with small finite samples, ambiguous graphical criteria, mixed observational-interventional data, and counterfactual estimands. Code is available at https://github.com/lbynum/metadentify.

Key takeaway

For causal inference practitioners evaluating the feasibility of estimating effects with finite, noisy data, this framework provides an actionable approach. You should consider using computational identifiability to empirically determine if a target parameter is estimable within a desired error tolerance, rather than relying solely on theoretical guarantees. This method helps in selecting optimal adjustment sets, assessing transportability with mixed data, and understanding the practical limits of CATE versus ITE estimation.

Key insights

Computational identifiability reframes identification as a finite, empirical search for an estimator within specified error bounds.

Principles

• Identifiability can be conditional on a prior over SCMs and a hypothesis space.
• Approximation within an ε margin is sufficient, not perfect prediction.
• Computational identifiability can have non-monotonic relationships with dataset size.

Method

Meta-learn an estimator φ from observations and query points to causal query values, using a joint causal-query-mixture distribution.

In practice

• Distinguish optimal adjustment sets when theoretical criteria are ambiguous.
• Evaluate identifiability with mixed observational-interventional data for transportability.
• Compare CATE vs. ITE identification under finite samples and noise.

Topics

• Computational Identifiability
• Causal Inference
• Meta-learning
• Structural Causal Models
• Causal Effect Estimation
• Finite Sample Analysis

Code references

• lbynum/metadentify

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Identifiability of Causal Graphs under Non-Additive Conditionally Parametric Causal Models — CPCMs offer flexible causal discovery by modeling conditional distributions with diverse parametric families.
• Identifying Weight-Variant Latent Causal Models — Identifying latent causal variables requires addressing transitivity, permutation, and scaling indeterminacies, especially via modulated linear-Gaussian models.
• Optimal Experiments for Partial Causal Effect Identification — Optimal experimental design for partially identifiable causal queries is NP-hard but can be significantly pruned using graphical criteria.
• Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions — Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.
• Identifying Information from Observations with Uncertainty and Novelty — Identification and sample complexity are unified across computational, statistical, and PAC-Bayes learning frameworks.
• Clustering and Pruning in Causal Data Fusion — Pruning and clustering reduce causal graph complexity in multi-source data fusion while preserving causal effect identifiability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.