Abstract
Weighting estimators based on propensity scores are widely used for causal estimation in a variety of contexts, such as observational studies, marginal structural models and interference. They enjoy appealing theoretical properties such as consistency and possible efficiency under correct model specification. However, this theoretical appeal may be diminished in practice by sensitivity to misspecification of the propensity score model. To improve on this, we borrow an idea from an alternative approach to causal effect estimation in observational studies, namely subclassification estimators. It is well known that compared to weighting estimators, subclassification methods are usually more robust to model misspecification. In this paper, we first discuss an intrinsic connection between the seemingly unrelated weighting and subclassification estimators, and then use this connection to construct robust propensity score weights via subclassification. We illustrate this idea by proposing so-called full-classification weights and accompanying estimators for causal effect estimation in observational studies. Our novel estimators are both consistent and robust to model misspecification, thereby combining the strengths of traditional weighting and subclassification estimators for causal effect estimation from observational studies. Numerical studies show that the proposed estimators perform favorably compared to existing methods.
Introduction
In observational studies, the propensity score—the probability of receiving a treatment given a set of observed covariates—serves as a cornerstone for mimicking the conditions of a randomized controlled trial. Traditionally, researchers have relied on two primary techniques to utilize these scores: Inverse Probability Weighting (IPW) and Subclassification.
While IPW is theoretically efficient and consistent under correct model specification, it is notoriously sensitive to model misspecification. When the estimated propensity scores are near zero or one, weights can "explode," leading to extreme variance and biased estimates. Conversely, subclassification—grouping units into a small number of bins (typically five)—is prized for its robustness because it relies on the rank of the scores rather than their exact values. However, subclassification is often criticized for being inconsistent, as a fixed number of bins usually fails to fully remove confounding bias. This paper introduces a "full-classification" weighting scheme that bridges this gap, offering the robustness of subclassification with the mathematical consistency of weighting.
The Methodology of Full Subclassification (FS)
The authors propose Full Subclassification (FS) weights, a method that utilizes model-based scores only to rank and bin observations. The process follows three distinct steps:
- Rank and Cluster: A standard parametric model (such as a logistic regression) is used to estimate propensity scores. However, these scores are used solely to sort and group units into $K$ subclasses.
- Empirical Weighting: Instead of using the predicted probabilities from the model, the weight for each bin is calculated based on the actual, empirical proportion of treated versus control units found within that specific bin.
- Dynamic Binning: Unlike traditional subclassification, which uses a fixed number of bins, the FS approach allows the number of subclasses $K$ to grow as the sample size $N$ increases.
The "Full" classification recommendation suggests choosing the maximum number of subclasses ($K_{max}$) such that every bin contains at least one treated and one control unit. The authors demonstrate that this approach achieves root-N consistency, converging to the true effect at the same rate as standard weighting methods while maintaining significantly higher stability.
Theoretical Insights: Unifying Estimators
A significant theoretical contribution of this work is Proposition 1, which reveals an intrinsic connection between seemingly disparate estimators. The authors show that when using FS weights, the Horvitz-Thompson estimator and the Ratio estimator become mathematically identical. This unification simplifies the decision-making process for practitioners and provides a more stable foundation for causal inference.
Furthermore, the FS method addresses the "Stability-Balance Trade-off." Modern methods like Covariate Balancing Propensity Scores (CBPS) often force perfect covariate balance at the cost of extreme or negative weights. FS weights are naturally "coarsened," ensuring they remain within a reasonable range (typically $[0, 1]$). This provides a "Goldilocks" solution: sufficient balance between groups without the instability of extreme weights.
Empirical Performance
Simulation Studies
Using the complex simulation framework established by Kang and Schafer (2007)—which involves non-linear transformations of covariates to intentionally misspecify the propensity score model—the authors tested the FS estimator's resilience. While the bias of standard Ratio estimators exploded as the sample size grew under misspecification, the FS estimator remained stable and accurate, demonstrating superior robustness.
Application: NHANES Childhood Nutrition Study
The method was applied to real-world data from the 2007–2008 National Health and Nutrition Examination Survey (NHANES), involving 2,330 children. The goal was to study the impact of school meal programs on Body Mass Index (BMI).
Traditional weighting methods produced highly unstable weights, ranging from -614 to 14,940. In contrast, the FS weights remained tightly clustered. This stability led to more reliable confidence intervals and a clearer causal finding: participation in school meal programs had a negligible effect on BMI once confounding was robustly controlled. Notably, the FS method yielded nearly identical results regardless of whether a logistic or a complementary log-log model was used, proving its relative indifference to the initial model's functional form.
Conclusion and Future Directions
The Full Subclassification weighting scheme provides a robust alternative for causal inference in observational studies. By transforming the propensity score model into a ranking tool rather than a direct source of weights, it shields the estimator from the volatility of misspecified models.
Future research may focus on optimizing the growth rate of $K$ relative to $N$ for small-to-medium samples, exploring more sophisticated clustering techniques beyond simple quantiles, and extending the FS framework to multi-arm treatments and generalized propensity scores.
Summary updated