A Sensitivity Approach to Causal Inference Under Limited Overlap

1. Problem Statement

In observational causal inference, a critical challenge arises when there is limited overlap (or lack of common support) between the treatment and control groups. This occurs when the propensity score $\pi(x) = P(Z=1|X=x)$ is close to 0 or 1 for certain regions of the covariate space.

• Consequences: Standard estimators (e.g., Inverse Probability Weighting - IPW, Augmented IPW - AIPW) suffer from high variance or infinite variance in these regions.
• Current Practices & Limitations: To stabilize estimates, researchers often employ trimming (discarding units with extreme propensity scores) or reweighting. While this reduces variance, it fundamentally changes the estimand (shifting focus to the "overlap" region) and introduces bias in estimating the overall Average Treatment Effect (ATE).
• The Gap: Standard asymptotic confidence intervals often fail silently in these scenarios, providing misleadingly narrow intervals that do not cover the true effect. Existing minimax approaches that attempt to bound this bias are often overly conservative, resulting in intervals so wide they are practically useless.

2. Methodology

The authors propose a sensitivity framework that combines standard asymptotic inference with finite-sample minimax inference to quantify and bound the bias introduced by trimming.

Core Framework: Decomposition and Minimax

The method decomposes the Average Treatment Effect (ATE), $\tau$, into two components based on a user-defined overlap threshold $\epsilon$:

1. $\tau_+$ (Overlap Region): The ATE for units where the propensity score is within a safe range ($q(x) \ge \epsilon$). This is estimated using standard asymptotic methods (e.g., AIPW), assuming the sample size is sufficient.
2. $\tau_-$ (Non-Overlap Region): The ATE for units with limited overlap ($q(x) < \epsilon$). This is the source of bias when trimming is applied.

The authors apply a minimax approach specifically to $\tau_-$ rather than the entire dataset.

• Function Class: They assume the outcome function $f(x, z)$ belongs to a class of Lipschitz continuous functions with constant $L$.
• Modulus of Continuity: Using Donoho's minimax framework, they compute the modulus of continuity $\omega(\delta)$ to find the tightest possible confidence interval that accounts for the worst-case bias across all admissible outcome functions.
• Estimator Construction: The resulting estimator, Minimax Partial (MP), provides a worst-case confidence bound for the bias introduced by ignoring the non-overlap region.

Sensitivity Analysis via Contextualization

A key innovation is avoiding the need for the analyst to arbitrarily commit to a specific Lipschitz constant $L$ (which is often uninterpretable). Instead, they propose a data-driven contextualization:

• Procedure: Estimate the smoothness of the outcome function within the overlap region (where data is abundant).
• Metric: Calculate the $p$-th quantile of the local Lipschitz constants derived from the fitted outcome regressor in the overlap region.
• Application: Use this data-derived $L$ (or a range of percentiles) to construct the minimax bounds for the non-overlap region. This translates abstract smoothness assumptions into interpretable diagnostics.

Hybrid Confidence Intervals

To provide a valid inference for the total ATE, the authors propose a combined interval ($MP_{combine}$):
$$MP_{combine} = AIPW_{partial} + MP$$
Using a union bound, they combine the asymptotic interval for the overlap region with the minimax bias bound for the non-overlap region. This yields a valid confidence interval for the full ATE that is significantly narrower than applying minimax inference to the whole dataset.

3. Key Contributions

1. Targeted Minimax Inference: Unlike previous works that apply minimax bounds to the entire dataset (leading to excessive conservativeness), this method isolates the non-overlap region. This allows standard asymptotic methods to be used where they are valid, while only resorting to worst-case bounds where necessary.
2. Sensitivity Framework: The paper introduces a systematic way to assess how "irregular" the outcome function must be to invalidate a finding. By varying the smoothness parameter $L$ (via percentiles), analysts can visualize the trade-off between bias and interval width.
3. Data-Driven Contextualization: The method bridges the gap between theoretical minimax bounds and practical application by calibrating the Lipschitz constant using observed data from the overlap region.
4. Guidance for Data Collection: The framework is extended to active data collection settings. It provides a metric to evaluate which new data points to sample to most effectively reduce uncertainty in the non-overlap region, outperforming simple propensity-score-based heuristics.

4. Results

The authors validate their approach through simulations and a real-world case study (PennUI dataset).

• Simulation (Example 1):
  • Standard AIPW with trimming ($AIPW_{partial}$) produced confidence intervals that failed to cover the true ATE as the non-overlap region increased.
  • The proposed MP interval successfully covered the bias in the non-overlap region.
  • The combined interval ($MP_{combine}$) achieved valid coverage for the total ATE with significantly shorter length compared to the naive full-minimax approach.
• Real Data (PennUI):
  • Applied to a reemployment bonus experiment. As the degree of non-overlap was artificially increased, $AIPW_{partial}$ showed high under-coverage.
  • The sensitivity analysis revealed that the bias was highly sensitive to the assumed smoothness $L$.
  • The method successfully identified that the bias was non-negligible under realistic smoothness assumptions, whereas standard methods suggested the results were robust.
• Data Collection:
  • In a simulated active learning scenario, the metric derived from the MP framework correctly identified that sampling data in "hard-to-extrapolate" regions (Option 2) was more valuable than sampling in "easy-to-extrapolate" regions (Option 1), even when propensity scores looked similar.

5. Significance

This paper addresses a fundamental limitation in observational causal inference: the silent failure of standard methods under limited overlap.

• Practical Utility: It provides a diagnostic tool for practitioners to determine if their causal conclusions are driven by reliable data or by dangerous extrapolation.
• Balanced Inference: It offers a middle ground between the "naive" approach (ignoring bias) and the "overly conservative" approach (giving up on inference).
• Interpretability: By linking the Lipschitz constant to data-driven percentiles, it makes the sensitivity analysis accessible to domain experts who may not be familiar with minimax theory.
• Future Directions: The framework lays the groundwork for adaptive data collection and off-policy evaluation in high-dimensional settings, ensuring that decisions are made with a quantified understanding of uncertainty.

In summary, the authors provide a rigorous, practical, and interpretable framework to quantify the uncertainty of causal effects in regions where data is sparse, preventing spurious findings and guiding more robust policy decisions.