For decades, the field of causal inference has been divided into two distinct tasks: identification and estimation.
Identification is a purely symbolic exercise. It asks a foundational "what-if" question: If we had an infinite amount of observational data, could we mathematically isolate the true causal effect of an intervention? Using frameworks like Judea Pearl’s do-calculus or the ID algorithm developed by Ilya Shpitser et al. in Identification of Joint Interventional Distributions in Recursive Semi-Markovian Causal Models, researchers proved that many complex causal queries could be reduced to symbolic formulas containing only observed probabilities.
Estimation, on the other hand, is statistical. It asks a practical question: How do we actually compute this effect using a finite, noisy, and often high-dimensional dataset?
For a long time, literature treated identification as a solved problem and estimation as a separate implementation detail. However, a major realization has emerged across the field: the identification-estimation gap. A causal effect can be perfectly identifiable in theory, yet completely unestimable in practice using naive statistical methods. Bridging this gap requires rethinking how we translate abstract causal graphs into robust statistical algorithms.
Why Identification Formulas Fail in the Finite World
To understand why the identification-estimation gap exists, we must look at what happens when an abstract mathematical formula meets real-world data constraints.
1. The Positivity and Support Trap
Theoretical identification algorithms operate on population-level distributions, assuming we know the exact probability of every possible event. In practice, we work with finite samples. As Yaroslav Kivva et al. demonstrate in Revisiting the general identifiability problem, standard causal reasoning tools can completely fail if the "positivity assumption"—the requirement that all possible observations have a non-zero probability—is violated or poorly estimated in finite samples. If a specific sub-population rarely receives a treatment, the denominator in an identification formula approaches zero, causing statistical variance to skyrocket.
2. Numerical Instability and Error Amplification
Identification formulas often involve complex, nested chains of conditional probabilities or matrix inversions. Leonard J. Schulman et al. show in Stability of Causal Inference that these symbolic calculations can be dangerously unstable. In certain graph structures, tiny estimation errors in the input data can explode exponentially throughout the formula. A researcher might require an impossible level of data precision just to get a meaningful, non-random causal estimate.
3. The Curse of Dimensionality
When causal paths become intricate—such as the path-specific or personalized effects studied by Daniel Malinsky et al. in A Potential Outcomes Calculus for Identifying Conditional Path-Specific Effects—the resulting identification formulas require estimating high-dimensional conditional distributions. Traditional non-parametric "plug-in" estimators (which simply replace true probabilities with empirical frequencies) break down rapidly as the number of covariates grows.
The Solution: Double Machine Learning for General Graphs
The most successful strategy for closing the identification-estimation gap is to automatically derive efficient influence functions (EIFs) for complex causal expressions. This enables the deployment of Double Machine Learning (DML).
In traditional machine learning, if you use a flexible model (like a neural network) to estimate a conditional probability within a causal formula, the model's inherent biases—caused by regularization or overfitting—will leak directly into your causal estimate.
DML bypasses this via Neyman-orthogonality. Instead of plugging data directly into an identification formula, DML constructs a "score function" where the first-order derivative with respect to the nuisance parameters (the intermediate machine learning models) is exactly zero. In intuitive terms, DML subtracts out the predictable components of both the treatment and the outcome, isolating the clean causal signal. This ensures that errors in the machine learning models only have a negligible, second-order impact on the final causal estimate, restoring root-$n$ statistical consistency.
Naive Plug-in: Data ──> [ Flexible ML Models ] ──> [ Identification Formula ] ──> Biased Estimate
DML Approach: Data ──> [ Flexible ML Models ] ──┐
[ Orthogonal Score ] <─┴───────────────────────────────> Unbiased Estimate
While DML was originally restricted to simple backdoor adjustments, Jin Tian et al. developed a general framework called DML-ID in Estimating Identifiable Causal Effects through Double Machine Learning. Their algorithm takes any arbitrary, identifiable causal graph, automatically computes its symbolic identification formula, derives its corresponding influence function, and outputs a robust DML estimator.
The authors later extended this paradigm to handle structural uncertainty. In Estimating Identifiable Causal Effects on Markov Equivalence Class through Double Machine Learning, the authors proved that you do not even need a perfectly specified causal graph; you can derive robust DML estimators directly from a Markov Equivalence Class (represented as a Partial Ancestral Graph), bridging the gap between structure learning and robust estimation.
Alternative Bridges: Weighting Operators and ERM
Beyond DML, researchers have developed other structural transformations to make identification formulas statistically well-behaved.
Weighting Operators
Rather than evaluating a complex graph all at once, Jin Tian et al. introduced the concept of "weighting operators" in Estimating Causal Effects Using Weighting-Based Estimators. This technique decomposes advanced graphical criteria (like the front-door criterion or surrogate endpoints) into a sequential chain of reweighted regressions. By mapping abstract graph operations to generalized Inverse Probability Weighting, they provide a familiar, stable numerical path for estimation.
Weighted Empirical Risk Minimization
In Learning Causal Effects via Weighted Empirical Risk Minimization, the authors bypass plug-in estimation entirely. They transform the symbolic identification formula into a loss function, allowing a model to learn causal effects by directly minimizing a weighted empirical risk over finite data.
The Gap in Complex Data Regimes
The necessity of bridging identification and estimation becomes glaringly obvious when moving to advanced causal paradigms:
- Multi-Source Data (General Identifiability): Theoretical frameworks like g-identifiability, detailed by Sanghack Lee in General Identifiability with Arbitrary Surrogate Experiments, prove when you can combine multiple observational and experimental datasets to identify an effect. To make this useful for finite datasets, Yonghan Jung et al. introduced the MR-gID estimator in Estimating Causal Effects Identifiable from a Combination of Observations and Experiments, providing the statistical machinery needed to safely merge these data streams without exploding variance.
- Partial Identification and Bounds: When an effect is unidentifiable, researchers calculate mathematical bounds. However, Alexis Bellot notes in Towards Bounding Causal Effects under Markov Equivalence that standard bounding frameworks assume perfect population distributions. In reality, finite-sample sampling noise introduces massive statistical uncertainty into the boundaries themselves, meaning the estimated intervals must be regularized and corrected for finite-sample variance (Noam Finkelstein et al. in Partial Identifiability in Discrete Data with Measurement Error).
By treating identification and estimation as an interconnected pipeline rather than isolated problems, modern causal inference ensures that if a causal question can be answered on paper, it can be computed reliably from real data.
Key Foundational Work
Identification of Joint Interventional Distributions in Recursive Semi-Markovian Causal Models (2006) — contributed 1 concept elements to this direction.
A New Characterization of the Experimental Implications of Causal Bayesian Networks (2002) — contributed 1 concept elements to this direction.