Retrospective Counterfactual Prediction by Conditioning on the Factual Outcome: A Cross-World Approach

1. Problem Statement

The paper addresses the challenge of retrospective counterfactual prediction. In causal inference, this involves estimating the potential outcome $Y(1)$ for an individual who actually received treatment $T=0$ (observed outcome $Y(0)$), given their covariates $X=x$.

• The Target Quantity: The goal is to estimate the conditional expectation $\mu(x, y) = E[Y(1) \mid X=x, Y(0)=y]$ and construct valid prediction intervals $C(x, y)$ such that $P(Y(1) \in C(x, y) \mid X=x, Y(0)=y) \geq 1-\alpha$.
• The Core Challenge: Unlike estimating Conditional Average Treatment Effects (CATE), which relies only on covariates $X$, estimating individual counterfactuals conditioned on the observed outcome $Y(0)$ is fundamentally unidentifiable from observational data alone. The joint distribution of potential outcomes $(Y(0), Y(1))$ is never observed for the same unit.
• Limitations of Existing Methods:
  • Direct Outcome Modeling: Estimates $E[Y(1) \mid X]$ ignoring $Y(0)$, losing individual-specific information.
  • CATE-Adjusted Imputation: Uses $\hat{Y}(1) = Y(0) + \hat{\tau}(X)$. This implicitly assumes a specific dependence structure (often equivalent to $\rho=1$ or constant effects) and ignores uncertainty in the treatment effect estimation.
  • Matching: Implicitly assumes independence between potential outcomes ($\rho=0$).
  • Bayesian/Adversarial Methods: Often rely on strong, untestable modeling assumptions or black-box architectures without explicit control over the cross-world dependence.

2. Methodology: The Cross-World Approach

The author proposes a framework called Retrospective Counterfactual Prediction (RCP) that explicitly models the dependence between potential outcomes using a cross-world correlation parameter $\rho(x)$.

A. The Cross-World Assumption

The method relies on the parameter $\rho(x) = \text{cor}(Y(1), Y(0) \mid X=x)$.

• Identifiability: Lemma 2 proves that without specifying $\rho(x)$, the quantity $\mu(x, y)$ is not identified. Different dependence structures can yield identical observed data but different counterfactual predictions.
• Role of $\rho(x)$: It acts as a sensitivity parameter. While not identifiable from data alone, it can be guided by domain knowledge, auxiliary data (e.g., genetic markers in a subset of patients), or sensitivity analysis.
• Gaussian Copula: For tractability, the paper assumes a Gaussian copula structure linking $Y(0)$ and $Y(1)$, fully specified by $\rho(x)$.

B. The Estimator Construction

The framework combines baseline estimates for the treated and control groups with the factual outcome $y$:

1. Baseline Estimates: Obtain point estimates $\hat{\mu}_t(x)$ and prediction intervals $\tilde{C}_t(x)$ for $t \in \{0, 1\}$ using standard methods (e.g., Random Forests, Quantile Regression).
2. Point Estimator ($\hat{\mu}_\rho$):
   $$\hat{\mu}_\rho(x, y) = \hat{\mu}_1(x) + \rho \cdot \lambda(x) \cdot (y - \hat{\mu}_0(x))$$
   Where $\lambda(x)$ adjusts for the relative width (variance) of the baseline prediction intervals.
   • If $\rho=0$, the estimate ignores $y$ (reverts to $\hat{\mu}_1(x)$).
   • If $\rho=1$, the estimate becomes $y + \hat{\tau}(x)$ (perfect correlation).
3. Prediction Intervals ($C_\rho$):
   The interval width is shrunk based on the information gained from conditioning on $Y(0)$. The width scales by $\sqrt{1-\rho^2}$.
   $$C_\rho(x, y) = \left[ \hat{\mu}_\rho(x, y) - \sqrt{1-\rho^2} \cdot l_{1-T}(x), \quad \hat{\mu}_\rho(x, y) + \sqrt{1-\rho^2} \cdot u_{1-T}(x) \right]$$

C. Finite-Sample Bias Correction ($C^{+CI}_\rho$)

Since $\hat{\mu}_\rho$ is an estimator, it introduces estimation error. For high $\rho$, this error can lead to undercoverage. The paper introduces a bias-corrected interval that adds a scaled confidence interval (derived via bootstrapping) around the point estimate:
$$C^{+CI}_\rho(x, y) = \hat{\mu}_\rho(x, y) \pm \left( c \cdot r(x,y) + \sqrt{1-\rho^2} \cdot \text{width} \right)$$
The scaling factor $c = \rho^2$ ensures a smooth interpolation between no correction ($\rho=0$) and full correction ($\rho=1$).

3. Key Contributions

1. Formalization of Cross-World Dependence: The paper provides the first systematic treatment of retrospective counterfactual prediction by explicitly parameterizing the dependence between potential outcomes via $\rho(x)$.
2. Theoretical Guarantees: Under the assumption of a correctly specified Gaussian cross-world dependence, the proposed estimators are:
   • Consistent: Converge to the true conditional mean.
   • Asymptotically Optimal: The prediction intervals achieve the smallest possible width for a given coverage level.
   • Valid: Asymptotically satisfy the coverage probability $1-\alpha$.
3. Bridging Extremes: The method unifies existing approaches as special cases:
   • $\rho=0$: Equivalent to ignoring the factual outcome (Direct Outcome modeling).
   • $\rho=1$: Equivalent to assuming constant treatment effects (CATE-adjusted imputation).
   • $0 < \rho < 1$: Interpolates between these, leveraging the factual outcome to reduce variance.
4. Empirical Validation: Extensive experiments on synthetic, semi-synthetic (IHDP), and real-world (Twins) datasets demonstrate that conditioning on $Y(0)$ significantly improves accuracy (lower MSE) and interval efficiency (better Interval Scores) when $\rho > 0$.

4. Experimental Results

• Datasets: Synthetic data (varying dimensions and $\rho$), IHDP (semi-synthetic), and Twins (real-world).
• Baselines: Compared against Direct Outcome (DO), CATE-adjusted imputation (T-learner), Matching, and GANITE.
• Performance:
  • Point Estimation: RCP($\rho$) consistently outperforms baselines in Mean Squared Error (MSE) when $0 < \rho < 1$. The improvement scales with $\rho$, following the theoretical variance reduction factor of $(1-\rho^2)$.
  • Prediction Intervals: The proposed intervals achieve better Interval Scores (balancing width and coverage) than baselines.
  • Robustness: The method remains effective even when $\rho$ is misspecified (e.g., using $\rho_{est} \approx \rho_{true}$), provided the misspecification is not extreme.
  • Non-Gaussianity: The method is robust to non-Gaussian marginal distributions, though convergence is slightly slower in finite samples with heavy tails.
• Recommendation: Use uncorrected intervals $C_\rho$ when $\rho \leq 0.5$ and bias-corrected intervals $C^{+CI}_\rho$ when $\rho > 0.5$.

5. Significance and Implications

• Practical Utility: In high-stakes domains like medicine (e.g., "Would this patient have survived if treated?"), ignoring the observed outcome $Y(0)$ discards critical individual-level information. This method provides a principled way to incorporate that information.
• Transparency: By making the cross-world dependence assumption explicit via $\rho$, the method forces practitioners to confront the untestable assumptions required for individual-level inference. It shifts the burden from "black-box" modeling to "sensitivity analysis" regarding $\rho$.
• Future Directions: The paper suggests extending this framework to continuous treatments, dynamic settings (time series), and developing tools to better estimate or bound $\rho$ using auxiliary data.

In summary, this paper advances causal inference by moving beyond average effects to precise individual counterfactuals, demonstrating that conditioning on the factual outcome is not only possible but highly beneficial, provided one explicitly models the cross-world correlation.