Estimation and Inference for the Average Treatment Effect in a Score-Explained Heterogeneous Treatment Effect Model

Kevin Christian Wibisono, Debarghya Mukherjee, Moulinath Banerjee, Ya’acov RitovarXiv 2025 · DOI · arXiv

Abstract

In many practical situations, randomly assigning treatments to subjects is uncommon due to feasibility constraints. For example, economic aid programs and merit-based scholarships are often restricted to those meeting specific income or exam score thresholds. In these scenarios, traditional approaches to estimating treatment effects typically focus solely on observations near the cutoff point, thereby excluding a significant portion of the sample and potentially leading to information loss. Moreover, these methods generally achieve a non-parametric convergence rate. In this study, we propose a differencing and matching-based estimator of the average treatment effect on the treated (ATT) in the presence of heterogeneous treatment effects, utilizing all available observations. We establish the asymptotic normality of our estimator and demonstrate its effectiveness through various synthetic and real data analyses.

Introduction: Beyond Local Effects

In causal inference, the randomized controlled trial (RCT) is often considered the "gold standard." However, in many real-world policy and economic contexts, treatment is assigned based on a score or threshold. A student might receive a scholarship only if their test score exceeds a certain value, or a patient might receive a specific treatment only if their blood pressure reaches a clinical threshold.

Traditionally, researchers analyze these scenarios using Regression Discontinuity Design (RDD). While RDD is intuitive—comparing individuals just above and below the cutoff—it suffers from two major limitations:

  1. Information Loss: It discards data far from the threshold, focusing only on a narrow "local" window.
  2. Slow Convergence: Because it relies on a small subset of data, estimates improve at a non-parametric rate, which is slower than the standard $\sqrt{n}$ rate seen in parametric models.

Wibisono et al. address these issues by introducing a global approach that utilizes the entire dataset while accounting for heterogeneous treatment effects—the reality that a treatment may affect different individuals in different ways.

Methodology: A Score-Explained Framework

The core of the proposed method is a multi-step algorithm that combines linear regression, non-parametric differencing, and residual matching. The model assumes that the treatment assignment is "score-explained," meaning the score variable ($Q$) is the primary driver of treatment, but it may be influenced by unobserved confounders ($\eta$).

Causal structure of the score-explained model, showing the relationships between the unobserved confounder (η), the score (Q), covariates (X), and the outcome (Y).

The estimation process follows these key steps:

  1. Isolating the "Innate" Factor: The researchers model the score $Q$ as a function of observed covariates $X$ and an unobserved error term $\eta$. This $\eta$ represents latent factors, such as "innate ability" or "grit," that influence both the score and the final outcome.
  2. First-Order Differencing: To handle the unknown non-parametric functions in the model, the algorithm sorts the control group by their estimated innate factors ($\hat{\eta}$) and takes differences between adjacent observations. This technique "cancels out" complex, unknown baseline effects.
  3. Residual Matching: For each treated individual, the algorithm identifies a "nearest neighbor" in the control group with a similar estimated innate factor. By comparing their outcomes, the individual treatment effect can be isolated.
  4. Cross-Fitting: To ensure statistical robustness and prevent overfitting, the authors employ data splitting, where parameters are estimated on one portion of the data and applied to another.

Key Insights and Results

The primary contribution of this work is proving that the proposed estimator achieves a parametric convergence rate ($\sqrt{n}$). This means that even though the model allows for complex, non-parametric relationships and heterogeneous effects, the estimates improve as quickly as a standard linear regression as the sample size increases.

Real-World Applications

The authors validated their approach using several datasets, including:

  • Academic Probation: Analyzing data from Lindo et al. (2010), the study examined how being placed on academic probation (based on a GPA cutoff) affects future performance. The model found an ATT of 0.28, suggesting that probation acts as a "wake-up call" that improves subsequent grades.
  • Women's Empowerment: Using data from Meyersson (2014), the researchers explored the impact of Islamic political party success on women's education in Turkey.

Beyond the average effect (ATT), the method also produces estimates for the Conditional Average Treatment Effect (CATE) and Individual Treatment Effect (ITE) as byproducts. Simulations showed that the predicted ITE curves closely mirrored the true underlying treatment effects, demonstrating the model's high fidelity.

Conclusion and Future Directions

This research provides a significant advancement for causal inference in non-randomized settings. By moving from local RDD to a global score-explained model, researchers can gain insights into how treatments affect populations far from the assignment threshold.

Future work remains in several areas. While the authors use bootstrapping for variance estimation in practice, a formal mathematical proof of the bootstrap's consistency in this specific framework is still needed. Additionally, while the model provides point estimates for individual treatment effects, developing formal inference methods (like confidence intervals) for these individual-level estimates is a promising next step.

Summary updated

Abstract

In many practical situations, randomly assigning treatments to subjects is uncommon due to feasibility constraints. For example, economic aid programs and merit-based scholarships are often restricted to those meeting specific income or exam score thresholds. In these scenarios, traditional approaches to estimating treatment effects typically focus solely on observations near the cutoff point, thereby excluding a significant portion of the sample and potentially leading to information loss. Moreover, these methods generally achieve a non-parametric convergence rate. While some approaches, e.g., Mukherjee et al. (2021), attempt to tackle these issues, they commonly assume that treatment effects are constant across individuals, an assumption that is often unrealistic in practice. In this study, we propose a differencing and matching-based estimator of the average treatment effect on the treated (ATT) in the presence of heterogeneous treatment effects, utilizing all available observations. We establish the asymptotic normality of our estimator and illustrate its effectiveness through various synthetic and real data analyses. Additionally, we demonstrate that our method yields non-parametric estimates of the conditional average treatment effect (CATE) and individual treatment effect (ITE) as a byproduct.
BibTeX
@misc{wibisono2025estimation,
  author = {Kevin Christian Wibisono and Debarghya Mukherjee and Moulinath Banerjee and Ya’acov Ritov},
  title = {{Estimation and Inference for the Average Treatment Effect in a Score-Explained Heterogeneous Treatment Effect Model}},
  year = {2025},
  eprint = {2504.17126},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.17126},
  howpublished = {arXiv},
}
Figure