In standard supervised machine learning, we train models to predict an observable outcome $Y$ from a set of features $X$. If we want to know how a model performs, we simply compare its predictions against the true, observed labels in a test set.
Causal inference introduces a fundamental complication. Instead of predicting what is, we want to predict what would be under different interventions. Specifically, we want to estimate the Individual Treatment Effect (ITE)—the difference between what would happen to an individual if they received a treatment ($Y(1)$) versus if they did not ($Y(0)$).
The core obstacle is that we can only ever observe one of these two potential outcomes for a single individual. If a patient takes a drug, we observe $Y(1)$; their outcome without the drug, $Y(0)$, becomes a counterfactual that is forever hidden from us. As Jeffrey M. Wooldridge outlines in Recent Developments in the Econometrics of Program Evaluation and Guido Imbens reviews in Nonparametric Estimation of Average Treatment Effects under Exogeneity: A Review, this unobservability transforms causal inference from a standard prediction task into a complex missing data problem.
The Counterfactual Validation Gap
Because counterfactual outcomes are never observed in real-world data, we face what is known as the counterfactual validation gap. In traditional machine learning, hyperparameters are tuned and models are selected using cross-validation on a validation set. In ITE estimation, however, we lack the ground-truth labels ($Y(1) - Y(0)$) required to compute standard error metrics like Root Mean Squared Error (RMSE).
As Fredrik D. Johansson et al. point out in Learning Representations for Counterfactual Inference, this makes it exceptionally difficult to know whether an algorithm is overfitting to the factual data at the expense of its counterfactual predictions. Uri Shalit et al. expand on this in Estimating individual treatment effect: generalization bounds and algorithms, noting that selecting optimal regularization weights becomes a major challenge without an observable validation target.
This unobservability also breaks standard active learning workflows. When collecting data is expensive, active learning algorithms typically query instances where the model's label uncertainty is highest. However, Tian Qin et al. show in Budgeted Heterogeneous Treatment Effect Estimation that because the true individual effect is never directly observed, traditional error-reduction strategies cannot be straightforwardly applied to guide data collection.
Furthermore, this missing data problem limits our ability to calculate non-linear population metrics. In Limits of Approximating the Median Treatment Effect, Raghavendra Addanki et al. demonstrate that observed data only provides the marginal distributions of treated and control outcomes, not their joint distribution. Because many different joint distributions can yield the exact same marginal distributions, quantities like the Median Treatment Effect are fundamentally inestimable without making highly restrictive assumptions.
Bridging the Gap via Balanced Representations
To circumvent the lack of counterfactual labels, researchers often reframe ITE estimation from observational data as a domain adaptation problem.
In an ideal randomized controlled trial, the distribution of features $X$ is identical between the treated and control groups. In observational data, however, selection bias is rampant: sicker patients are more likely to receive a treatment. If a model is trained directly on this biased data, it will confuse the effect of the treatment with the effect of the underlying sickness.
To solve this, Fredrik D. Johansson et al. proposed mapping the high-dimensional feature space into a lower-dimensional latent representation space. The objective function forces this latent space to satisfy two competing goals:
- Predictive Accuracy: The representation must retain enough information to accurately predict the observed factual outcomes.
- Distribution Balancing: The representation must minimize the distance (using metrics like Wasserstein distance or Maximum Mean Discrepancy) between the treated and control distributions.
Original Features (Biased) Latent Representation (Balanced)
[Treated Group] --> (Selection) --> [Treated] \ Forced to
> Overlap
[Control Group] --> (Selection) --> [Control] / Perfectly
By forcing the latent distributions to overlap, the model is tricked into treating the observational data as if it came from a randomized trial. This allows the model to generalize across the factual-counterfactual divide, estimating how a control subject would behave in the treated domain and vice versa. Liuyi Yao et al. extended this intuition in Representation Learning for Treatment Effect Estimation from Observational Data. by showing that preserving local similarity among neighboring data points during this balancing process further stabilizes counterfactual predictions.
Generative Imputation and Problem Reframing
Rather than just learning invariant representations, another prominent strategy is to directly fill in the missing data using generative modeling.
In GANITE: Estimation of Individualized Treatment Effects using Generative Adversarial Nets, Jinsung Yoon et al. introduced a two-stage framework that treats counterfactual estimation as an imputation task. A generator network observes the factual outcome and attempts to generate the corresponding counterfactual outcome. A discriminator then tries to determine which outcome was real and which was imputed. Once this generative component creates a proxy "complete" dataset containing both $Y(1)$ and $Y(0)$ for every individual, a secondary predictive model can be trained using standard supervised learning.
Alternatively, the fundamental challenge can be mitigated by changing the nature of the prediction task itself. Nathan Kallus argues in Classifying Treatment Responders Under Causal Effect Monotonicity that trying to regress the exact continuous difference ($Y(1) - Y(0)$) is unnecessarily difficult. By introducing a monotonicity assumption—the assumption that the treatment can help or have no effect, but never cause harm—the problem can be reframed as a binary classification task: predicting whether an individual is a "responder" or a "non-responder," bypassing the need for full counterfactual regression.
Embracing Uncertainty: Partial Identification and Bounds
When point-identifying an individual counterfactual is impossible due to data constraints or unobserved confounding, researchers shift from finding a single exact value to establishing rigorous mathematical boundaries.
Ang Li 0009 et al. demonstrate in Probabilities of Causation: Role of Observational Data that while individual-level counterfactual probabilities (such as the probability that a treatment was both necessary and sufficient for an outcome) are unidentifiable from observational data alone, combining observational data with experimental data allows researchers to calculate tight upper and lower bounds for these quantities.
To make these intervals as narrow and informative as possible, Sirui Lin introduces a data-driven approach in Tightening Causal Bounds via Covariate-Aware Optimal Transport. By using optimal transport to map individual covariate profiles across the treated and control groups, the framework constrains the space of valid joint distributions, yielding significantly narrower partial identification intervals without requiring unprovable structural assumptions.