Non-overlap Average Treatment Effect Bounds

Technical Summary: Non-overlap Average Treatment Effect Bounds

Problem Statement
The Average Treatment Effect (ATE), defined as the mean difference in potential outcomes under treatment versus control, is a canonical target in causal inference. Standard identification of the ATE relies on three assumptions: consistency, conditional exchangeability (no unmeasured confounding), and overlap (positivity). The overlap assumption requires that every unit has a non-zero probability of receiving either treatment or control ($0 < P(A=1|X) < 1$).

When overlap fails—either structurally (propensity scores are exactly 0 or 1 for some subpopulations) or practically (propensity scores are extremely close to 0 or 1, leading to sparse data in finite samples)—the ATE is not point-identifiable. The standard solution in such scenarios is to abandon the ATE in favor of alternative estimands, such as effects defined on a trimmed "overlap" subpopulation or via incremental propensity score interventions. However, these approaches alter the scientific question of interest and often yield less interpretable results. Furthermore, traditional point estimators for the ATE (e.g., doubly-robust estimators) can exhibit poor finite-sample performance, infinite variance, or non-normal sampling distributions when overlap is violated.

Methodology
The authors propose a novel approach to derive non-overlap bounds: partial identification bounds for the ATE that do not require the overlap assumption, provided the outcome variable is bounded (e.g., $Y \in [0, 1]$).

1. Identification via Propensity Score Thresholding:
   The population is partitioned into two subpopulations based on a propensity score threshold $c \in [0, 0.5]$:

   • Overlap Subpopulation: Units where $\pi(X) > c$ and $\pi(X) < 1-c$. Here, the treatment effect is point-identifiable under standard assumptions.
   • Non-overlap Subpopulation: Units where $\pi(X) \le c$ or $\pi(X) \ge 1-c$. Here, the treatment effect is not identifiable.

   The authors derive bounds by combining the point estimate of the effect in the overlap region with worst-case bounds for the non-overlap region. Specifically, the lower bound $L(c)$ subtracts the probability mass of the high-propensity non-overlap group, and the upper bound $U(c)$ adds the probability mass of the low-propensity non-overlap group. The width of these bounds (the "identification gap") is proportional to the size of the non-overlap subpopulation.

2. Smooth Approximations:
   The derived bounds are non-smooth functionals due to the use of indicator functions, which poses challenges for estimation and inference under nonparametric assumptions. To address this, the authors construct smooth approximations ($s_\ell$ and $s_g$) of the indicator functions. These approximations are designed to satisfy specific properties (bounding the indicator functions from below) to ensure the resulting smooth bounds still contain the true ATE. This allows the use of semiparametric efficiency theory.

3. Estimation and Inference (TMLE):
   The authors propose a Targeted Minimum Loss-Based Estimator (TMLE) for the smooth bounds. By deriving the Efficient Influence Function (EIF) for these smooth functionals, they construct estimators that are $\sqrt{n}$-consistent and asymptotically normal under nonparametric conditions, provided the nuisance parameters (propensity score and outcome regression) are estimated at sufficient rates ($n^{-1/4}$).

4. Uniform Inference via Multiplier Bootstrap:
   Since the optimal threshold $c$ and smoothing parameter $\gamma$ are unknown, the authors develop a multiplier bootstrap procedure. This method constructs uniform confidence sets across a grid of $(c, \gamma)$ combinations. Researchers can then report the narrowest valid interval from this grid, balancing the trade-off between the identification gap (smaller $c$) and estimation uncertainty/smoothing error (larger $c$ or $\gamma$).

Key Contributions
The paper makes three primary contributions:

1. Identification: Derivation of partial identification bounds for the ATE that rely only on consistency, conditional exchangeability, and outcome boundedness, explicitly relaxing the overlap assumption.
2. Estimation and Inference: Development of $\sqrt{n}$-rate estimators (TMLE) and a multiplier bootstrap procedure that yields uniformly valid confidence sets across a range of thresholds and smoothing parameters. This allows for valid inference even when the non-overlap subpopulation is small.
3. Theoretical and Empirical Comparison: A formal comparison of non-overlap bounds against traditional point estimators across different overlap regimes (strong, somewhat weak, and very weak overlap).

Results

• Theoretical Findings:
  • Under strong overlap, the non-overlap confidence intervals are asymptotically equivalent to traditional point estimator intervals.
  • Under very weak overlap (where the efficiency bound for the ATE is infinite and $\sqrt{n}$-estimation is impossible for point estimators), the non-overlap bounds remain estimable at $\sqrt{n}$-rates with valid confidence intervals.
  • Under somewhat weak overlap (finite but large efficiency bound), the non-overlap bounds can yield shorter confidence intervals than point estimators in finite samples, particularly near the phase transition where the overlap is marginal.
• Simulation Studies:
  • In scenarios with extreme practical overlap violations, non-overlap bounds demonstrated higher statistical power, shorter uncertainty widths, and better empirical coverage compared to doubly-robust point estimators, which often produced uninformative intervals spanning the entire parameter space.
  • In "average-case" scenarios without extreme violations, the non-overlap bounds remained competitive, often achieving narrower intervals than traditional methods.
• Real-World Application:
  • The method was applied to six observational datasets, including the Right Heart Catheterization (RHC) study. In the RHC dataset, where traditional point estimators produced non-informative intervals due to overlap violations, the non-overlap bounds provided a tight, informative interval ($-0.10, -0.03$) that excluded zero, suggesting a significant treatment effect.

Significance
The authors claim that this work offers a "no-regret" solution for ATE estimation in the presence of overlap violations. Unlike alternative approaches that change the target estimand (e.g., trimming to an overlap population), this method maintains the ATE as the primary object of inquiry. It provides valid statistical inference even when the ATE is not point-identifiable, without requiring a priori knowledge of the overlap population. The approach is particularly valuable in real-world settings where overlap violations are often limited to small subpopulations, allowing researchers to report tight, valid bounds that leverage the signal from the overlap region while accounting for the uncertainty in the non-overlap region. The paper emphasizes that while there is a trade-off between the identification gap and estimation variance, the proposed uniform inference framework allows researchers to navigate this trade-off optimally.