Conformalized Causal Inference

Estimates a valid confidence interval for the causal effect of an intervention on a single unit, not just the population average. This moves beyond the Average Treatment Effect (ATE) to provide personalized uncertainty bounds.

• Uses a nonconformity measure based on the residuals of a causal model (e.g., Causal Forest, T-learner).
• Guarantees that the true individual treatment effect falls within the interval with a user-specified probability.
• Enables precise, high-stakes decisions like personalized medicine dosing or targeted marketing interventions.

Provides a finite-sample, marginal coverage guarantee for the estimated causal effect without assuming a specific distribution for the noise or the outcomes. This is critical because causal inference often relies on unverifiable assumptions.

• The guarantee holds under the standard exchangeability assumption of conformal prediction, applied to the causal estimation residuals.
• Robust to heteroscedasticity and complex, non-Gaussian error distributions common in real-world causal data.
• Offers a model-agnostic safety net, validating the uncertainty output of any black-box causal estimator.

Directly quantifies the uncertainty in the unobserved counterfactual outcome—what would have happened to a treated unit had it not been treated. This is the fundamental challenge of causal inference.

• Constructs prediction intervals for the missing counterfactual, not just the observed outcome.
• Uses split conformal prediction to calibrate on a held-out dataset where both factual and counterfactual outcomes are known (e.g., from a randomized trial).
• Provides a rigorous statistical basis for answering "What if?" questions with measurable confidence.

Protects against overly confident causal conclusions when the underlying causal model is misspecified. The conformal calibration step corrects the intervals to maintain validity even if the point estimate is biased.

• If the causal model underestimates uncertainty, the conformal wrapper widens the intervals to restore coverage.
• If the model is overly conservative, the wrapper tightens the intervals for more informative predictions.
• This is a critical safeguard when applying causal models to observational data where the true data-generating process is unknown.

A common and powerful implementation combines Generalized Random Forests for causal estimation with a conformal calibration layer. The forest estimates heterogeneous treatment effects, and conformal prediction provides valid intervals.

• The Causal Forest outputs an estimate of the Conditional Average Treatment Effect (CATE) and a measure of its variance.
• Conformal prediction uses the out-of-bag residuals from the forest as a nonconformity score.
• The result is a powerful, non-linear estimator with statistically valid, personalized uncertainty bounds.

Extends conformal causal inference to assess how sensitive the validity of the intervals is to violations of the unconfoundedness assumption. This quantifies the impact of hidden variables.

• By bounding the influence of a potential unobserved confounder, one can compute adjusted conformal intervals that remain valid under a specified level of confounding.
• Provides a formal, quantitative framework for answering: "How strong would an unmeasured confounder need to be to invalidate my conclusions?"
• Transforms a binary assumption into a continuous, auditable risk assessment.

The causal estimand at the heart of personalized decision-making, defined as the difference between a unit's potential outcome under treatment and its potential outcome under control. Unlike the Average Treatment Effect (ATE), the ITE captures heterogeneous responses across a population.

• Fundamental Problem: Only one potential outcome is ever observed; the other is a counterfactual.
• Conformalized Goal: Produce a prediction interval for the ITE that has finite-sample coverage, even if the underlying causal model is misspecified.

The act of estimating what would have happened to a specific individual had they received a different intervention. This requires a model to reason about an alternate reality not present in the data.

• Relies on strong identification assumptions like unconfoundedness and positivity.
• Conformal prediction wraps these point predictions with a nonconformity measure to output a valid uncertainty set, acknowledging the inherent ambiguity in the counterfactual.

The core identification assumption stating that, conditional on observed covariates, treatment assignment is independent of the potential outcomes. Also known as conditional ignorability or selection on observables.

• Violation: If an unobserved confounder exists, causal estimates are biased.
• Conformal Sensitivity: Conformalized causal inference can be extended with sensitivity analysis to bound coverage under specific violations of this assumption.

The Rubin Causal Model (RCM) that formalizes causality by defining a pair of outcomes for each unit: one if treated, one if not. The causal effect is the comparison of these two latent states.

• Notation: Y(1) for treatment, Y(0) for control.
• Conformal Connection: The framework provides the structured missing data problem that conformal prediction must quantify, treating the unobserved potential outcome as the target of a prediction set.

Conditional Average Treatment Effect estimation focuses on how the effect of an intervention varies across subpopulations defined by features. It bridges the gap between the population-level ATE and the individual-level ITE.

• Often estimated with meta-learners (S-Learner, T-Learner, X-Learner) or causal trees.
• Conformalizing a CATE model yields group-conditional coverage, ensuring the uncertainty intervals are valid for specific demographic or behavioral segments.

A technique that wraps a standard quantile regression model with a conformal calibration step to correct the predicted quantile intervals. This yields finite-sample valid prediction intervals without assuming a specific error distribution.

• Causal Application: Used to directly model the conditional quantiles of the potential outcome distribution.
• The calibration step adjusts for any systematic overconfidence or underconfidence in the base causal model's uncertainty estimates.