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

Juraj BodikCLeaR 2026 · OpenReview

TL;DR

This paper introduces a novel approach to predict what would have happened to an individual if they had received a different treatment, by leveraging their actual observed outcome. By explicitly modeling the relationship between potential outcomes (the "cross-world correlation"), they achieve more accurate and reliable counterfactual predictions, outperforming existing methods that ignore this crucial information.

What Problem Does This Paper Address?

In causal inference, a key goal is to understand the effect of a treatment. This often involves estimating counterfactuals: what would have happened to an individual if they had received a different treatment than they actually did. For example, if a patient received a drug and recovered, we might want to know if they would have recovered even faster without the drug. This is particularly challenging because we can never observe both potential outcomes (treated and untreated) for the same individual.

Existing methods often focus on estimating average treatment effects or predicting counterfactuals based solely on covariates ($X$). However, they largely ignore the factual outcome ($Y(0)$ for someone who wasn't treated, or $Y(1)$ for someone who was). This is a missed opportunity, as the factual outcome often contains valuable information about the individual's underlying health or characteristics, which can strongly inform what their counterfactual outcome might have been. The core problem is that the relationship between the treated and untreated potential outcomes is unidentifiable from observational data alone, requiring additional assumptions.

Methods

The paper proposes a new framework called Retrospective Counterfactual Prediction (RCP). The key innovation is to explicitly model the cross-world correlation, denoted by $\rho(x) = cor(Y(1), Y(0) | X=x)$. This parameter quantifies how strongly the potential outcomes are related, given the covariates.

Instead of just predicting $E[Y(1) | X=x]$, RCP aims to predict $E[Y(1) | X=x, Y(0)=y]$, effectively conditioning on both covariates and the observed factual outcome. They achieve this by:

  1. Assuming a dependence structure: They adopt a Gaussian copula to link the potential outcomes, parameterized by $\rho(x)$. This allows them to translate a chosen value of $\rho$ into a fully specified joint distribution.
  2. Estimating counterfactuals: They develop point estimators ($\hat{\mu}_{\rho}(x,y)$) and prediction intervals ($C_{\rho}(x,y)$) that leverage the assumed $\rho$.
  3. Interpolating between extremes: The method interpolates between two common baselines:
  4. $\rho = 0$: This corresponds to assuming independence between potential outcomes, effectively ignoring the factual outcome.
  5. $\rho = 1$: This corresponds to assuming a constant treatment effect, where the counterfactual is simply the factual outcome plus a fixed effect. RCP allows for values of $\rho$ between 0 and 1, using the factual outcome to refine predictions.

Datasets

The authors evaluate their method on three types of datasets:

  • Synthetic datasets: These are generated with controlled properties, allowing them to vary dimensionality, propensity scores, treatment effect functions, and crucially, the cross-world correlation $\rho$.
  • IHDP (semi-synthetic): This dataset uses real covariates from a randomized trial, with simulated counterfactual outcomes. They also consider a version with correlated noise to better reflect real-world scenarios where potential outcomes are not independent.
  • Twins (real-world): This dataset uses U.S. twin birth records, where twins within a pair are treated as natural counterfactuals. The outcome is one-year mortality.

Key Insights and Ideas

  • The factual outcome is gold: Ignoring the observed outcome ($Y(0)$ or $Y(1)$) is a significant missed opportunity for improving counterfactual predictions. The factual outcome provides valuable individual-specific information.
  • Cross-world correlation is key: The relationship between potential outcomes ($\rho$) is not just a theoretical curiosity; it's a practical parameter that directly impacts prediction accuracy.
  • Interpolation is powerful: By allowing $\rho$ to vary between 0 (no dependence) and 1 (perfect dependence), the method can adapt to different scenarios and leverage the factual outcome effectively. This is a more principled approach than simply assuming $\rho=0$ or $\rho=1$.
  • Domain knowledge matters: While $\rho$ cannot be estimated from data alone, domain expertise or auxiliary data can inform plausible values or bounds for $\rho$, leading to better predictions. $\rho$ acts as a sensitivity parameter.
  • Non-Gaussianity is a finite-sample issue: Violations of the Gaussianity assumption primarily affect the speed of convergence in finite samples, not the asymptotic validity of the method.

Key Results

  • Improved MSE: RCP consistently outperforms baseline methods (Direct Outcome, CATE-adjusted imputation, matching) in terms of Mean Squared Error (MSE) for counterfactual predictions, especially when the true cross-world correlation $\rho$ is non-zero. The improvement is more dramatic in lower dimensions.
  • Accurate Prediction Intervals: The proposed prediction intervals ($C_{\rho}$) achieve asymptotically perfect conditional coverage and are shown to be asymptotically optimal.
  • Sensitivity to $\rho$: The MSE decreases as $\rho$ increases (up to 1), reflecting the increased informativeness of the factual outcome. The method's performance degrades when $\rho$ is misspecified, but even approximate knowledge of $\rho$ yields significant gains over ignoring it.
  • Bias Correction for High $\rho$: For high values of $\rho$, a bias-corrected version of the prediction intervals ($C_{\rho+CI}$) is recommended to maintain coverage.

Limitations and Future Work

  • Specifying $\rho$: The main limitation is the necessity of specifying the cross-world correlation $\rho$. This is an inherent challenge in causal inference, as the relationship between potential outcomes is unidentifiable. However, the paper argues that all methods implicitly make assumptions about $\rho$, and RCP makes this explicit, allowing for informed choices.
  • Non-Gaussian Dependence: Future work could explore richer, non-Gaussian dependence structures for potential outcomes.
  • Continuous/Dynamic Treatments: Extending the methodology to continuous treatments or dynamic settings (e.g., time series) is a promising direction.
  • Guidance for $\rho$ Selection: Developing tools or diagnostics to guide the selection of $\rho$ in practice would broaden the applicability of cross-world inference.

Summary updated

Abstract

Retrospective causal questions ask what would have happened to an observed individual had they received a different treatment. We study the problem of estimating $\mu(x,y)=\mathbb{E}[Y(1)\mid X=x,Y(0)=y]$, the expected counterfactual outcome for an individual with covariates $x$ and observed outcome $y$, and constructing valid prediction intervals under the Neyman-Rubin superpopulation model. This quantity is generally not identified without additional assumptions. To link the observed and unobserved potential outcomes, we work with a cross-world correlation $\rho(x)=cor(Y(1),Y(0)\mid X=x)$; plausible bounds on $\rho(x)$ enable a principled approach to this otherwise unidentified problem. We introduce retrospective counterfactual estimators $\hat{\mu}_{\rho}(x,y)$ and prediction intervals $C(x,y)$ that asymptotically satisfy $P[Y(1)\in C(x,y)\mid X=x, Y(0)=y]\ge1-\alpha$ under standard causal assumptions. Many common baselines implicitly correspond to endpoint choices $\rho=0$ or $\rho=1$ (ignoring the factual outcome or treating the counterfactual as a shifted factual outcome). Interpolating between these cases through cross-world dependence yields substantial gains in both theory and practice.
BibTeX
@inproceedings{bodik2026retrospective,
  author = {Juraj Bodik},
  title = {{Retrospective Counterfactual Prediction by Conditioning on the Factual Outcome: A Cross-World Approach}},
  booktitle = {CLeaR},
  year = {2026},
  url = {https://openreview.net/forum?id=xVYy96X1yq},
}
Figure