Cross-World Assumption and Refining Prediction Intervals for Individual Treatment Effects

1. Problem Statement

The paper addresses a critical gap in causal inference: the quantification of uncertainty for Individual Treatment Effects (ITE).

• The Challenge: While Average Treatment Effects (ATE) and Conditional Average Treatment Effects (CATE) are well-established, estimating the specific effect for a single unit ($Y(1) - Y(0)$) is fundamentally difficult because the "fundamental problem of causal inference" prevents observing both potential outcomes simultaneously.
• Limitations of Current Methods: Existing methods for constructing prediction intervals for ITEs often suffer from two extremes:
  1. Over-conservatism: Methods that ignore the dependence between potential outcomes (e.g., assuming independence or using naive union bounds) produce intervals that are excessively wide, rendering them useless for high-stakes decision-making (e.g., personalized medicine).
  2. Lack of Guarantees: Methods that attempt to narrow intervals often lack rigorous finite-sample coverage guarantees due to the unidentifiability of the joint distribution of potential outcomes.
• The Core Obstacle: The joint distribution of $(Y(1), Y(0))$ given covariates $X$ is unidentifiable from observational data. Specifically, the cross-world correlation $\rho(x) = \text{cor}(Y(1), Y(0) \mid X=x)$ cannot be estimated without additional assumptions.

2. Methodology: The Cross-World Assumption

The authors propose a framework that leverages a Cross-World Assumption regarding the correlation $\rho(x)$ between potential outcomes.

A. The Cross-World Parameter ($\rho$)

The method introduces $\rho(x)$ as a sensitivity parameter. While unidentifiable, the authors argue that domain experts can often provide plausible bounds or reasonable estimates for $\rho(x)$ based on:

• Latent Structure: If unobserved factors (e.g., genetic predisposition, baseline health) affect both treatment and control outcomes similarly, $\rho(x)$ is likely positive and potentially high.
• Auxiliary Data: Using a subset of data with richer covariates ($Z$) to derive data-driven lower bounds on $\rho(x)$ via the law of total covariance.
• Sensitivity Analysis: Treating $\rho$ as a tunable parameter to explore robustness.

B. The CW($\rho$) Prediction Intervals

The core contribution is a new class of prediction intervals, denoted CW($\rho$), which explicitly incorporates the correlation parameter to tighten the interval width.

• Construction:
  1. Construct separate valid prediction intervals for the treated outcome $Y(1)$ and control outcome $Y(0)$, denoted $C_1(x)$ and $C_0(x)$, using standard conformal inference (e.g., Conformalized Quantile Regression).
  2. Combine these intervals to form the ITE interval $C_{ITE}(x)$.
  3. The width of the combined interval is determined by a Correlation-Adjusted Euclidean Distance function:
     $$D_\rho(l, u) = \sqrt{l^2 + u^2 - 2\rho l u}$$
     where $l$ and $u$ are the lower and upper widths of the individual outcome intervals.
• Mechanism:
  • If $\rho = 1$ (perfect positive correlation), the uncertainty in the difference $Y(1) - Y(0)$ is minimized (the interval shrinks significantly).
  • If $\rho = -1$ (perfect negative correlation), the uncertainty is maximized (the interval expands to the naive sum of widths).
  • The method allows for $\rho$ to be a known value, a lower bound, or a misspecified estimate.

C. The CW+CI($\rho$) Variant

To address scenarios where the CATE estimator $\hat{\tau}(x)$ may be biased (a common issue in high dimensions), the authors propose CW+CI($\rho$).

• This variant augments the CW($\rho$) interval with a confidence interval for the CATE estimator (e.g., derived via bootstrapping).
• It dynamically adjusts the weight of the confidence interval based on $\rho$ (e.g., adding full confidence intervals when $\rho=1$ and none when $\rho=-1$), ensuring validity even when the point estimate is imperfect.

3. Key Contributions

1. Novel Framework: The first rigorous framework to establish finite-sample and asymptotic guarantees for ITE prediction intervals under explicit cross-world assumptions.
2. Theoretical Optimality:
   • Proved that under Gaussian assumptions and known $\rho$, the CW($\rho$) intervals are asymptotically optimal (i.e., they are the smallest possible intervals satisfying the coverage constraint).
   • Established finite-sample coverage guarantees even when $\rho$ is only bounded (e.g., $\rho \geq \rho_{min}$) or when the CATE estimator has bias.
3. Practical Tools for $\rho$:
   • Provided a "Practitioner Recipe" for reasoning about $\rho$ using domain knowledge and latent structure models.
   • Developed a method to estimate data-driven lower bounds on $\rho$ using auxiliary covariates ($Z$) via Lemma 1.
4. Robustness: Demonstrated that the method remains valid and effective even if $\rho$ is slightly misspecified or if the noise distribution is non-Gaussian.

4. Experimental Results

The authors evaluated their method on synthetic data and semi-synthetic benchmarks (IHDP, Twins, Cross-over studies).

• Coverage vs. Width:
  • Competing methods (e.g., Lei & Candès, DR-meta-learners) either failed to maintain coverage when $\rho \neq 0$ or produced intervals 3 to 5 times wider than CW($\rho$).
  • CW($\rho$) achieved valid coverage (meeting the $1-\alpha$ target) while reducing interval width to often less than 1/3 of competing methods.
• Robustness to Misspecification:
  • Even when the assumed $\rho$ was misspecified (e.g., using $\rho_{used} = \rho_{true} - 0.25$), the method maintained valid coverage, albeit with slightly wider intervals.
  • The CW+CI($\rho$) variant successfully handled high-dimensional settings and biased CATE estimators where the base CW($\rho$) might under-cover.
• Real-World Data: Analysis of cross-over studies and the Twins dataset suggested that in many real-world scenarios, $\rho$ is indeed positive and often substantial ($\rho > 0.5$), justifying the use of these assumptions.

5. Significance and Impact

• Bridging the Identifiability Gap: The paper successfully navigates the "unidentifiability" of ITEs by shifting the burden from "proving $\rho$" to "bounding $\rho$ using expert knowledge," a paradigm shift similar to sensitivity analysis in observational studies.
• High-Stakes Decision Making: By providing narrow, valid prediction intervals, the method enables reliable risk assessment for individual patients or policy subjects, which is crucial for personalized medicine and precision policy.
• Statistical Validity: Unlike heuristic approaches that narrow intervals without guarantees, this method provides rigorous mathematical proofs for coverage, making it suitable for regulatory or clinical applications.
• Future Directions: The work opens avenues for integrating cross-world assumptions into meta-learning, continuous treatment settings, and time-series causal inference.

In summary, this paper transforms the problem of ITE uncertainty quantification from an intractable "black box" into a solvable problem by leveraging the often-overlooked correlation between potential outcomes, resulting in prediction intervals that are both statistically rigorous and practically useful.