📊 statistics

Principal stratification with U-statistics under principal ignorability

This paper extends principal stratification theory under principal ignorability to accommodate nonlinear contrasts via principal generalized causal effect estimands, deriving efficient influence functions that enable the development of multiply robust and debiased machine learning estimators based on U-statistics.

Original authors: Xinyuan Chen, Fan Li

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinyuan Chen, Fan Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to judge the effectiveness of a new training program (like a job corps) on people's future earnings. You have a simple problem: Not everyone follows the rules. Some people assigned to the program skip it, and some people in the control group sneak into the program anyway.

Furthermore, there's a second hurdle: Some people drop out or change their status (like getting a job or not) before you can even measure their final earnings. If someone doesn't get a job, their "weekly earnings" are undefined.

This paper introduces a new, smarter way to measure success in these messy situations. Here is the breakdown using everyday analogies:

1. The Old Way vs. The New Way

The Old Way (Average Difference):
Traditionally, researchers ask: "On average, how many more dollars did the treatment group make?"

The Flaw: This is like measuring a race by looking at the average speed. If one runner trips and falls (an outlier) or if the track is muddy (skewed data), the average speed becomes misleading. It also struggles with "ordinal" data—like ranking runners 1st, 2nd, or 3rd—where the difference between 1st and 2nd isn't necessarily the same as 2nd and 3rd.

The New Way (The "Win" and "Probabilistic" Approach):
The authors suggest asking different questions:

The "Probabilistic Index" (The Coin Flip): Instead of asking "How much more money?", ask: "If I pick one person from the treatment group and one from the control group at random, what is the chance the treatment person did better?"
- Analogy: Imagine a coin flip. If the coin lands heads 55% of the time, you know the treatment has a slight edge, even if you don't know exactly how much money was made. This is robust against weird outliers (like one person making a million dollars).
The "Win Ratio" (The Tournament): For ranked outcomes (like job levels), ask: "How many times does the treatment person 'win' against the control person, compared to how many times they 'lose'?"
- Analogy: It's like a tennis match score. If Treatment wins 3 matches for every 1 loss, the ratio is 3:1. This is great for comparing things that are ranked but not measured by a ruler.

2. The "Principal Stratification" Filter

The paper uses a concept called Principal Stratification. Think of this as sorting people into invisible "clubs" based on how they would behave in both scenarios (treatment and control), even though we can only see one scenario at a time.

The "Always-Compliers" Club: People who would take the treatment no matter what.
The "Never-Takers" Club: People who would never take the treatment.
The "Protected" Club: People who only take the treatment if assigned to it.

The authors argue that to get a fair answer, you must compare apples to apples within these specific clubs. You can't compare a "Never-Taker" in the treatment group to a "Complier" in the control group; that's like comparing a professional athlete to a casual jogger.

3. The "Triple-Proof" Safety Net

The biggest challenge is that we don't know exactly who belongs to which invisible club. We have to guess using math models. If our guess is wrong, our results are garbage.

The authors developed a new mathematical tool (using U-statistics) that acts like a Triple-Proof Safety Net.

To get a correct answer, you usually need three different models to be perfect:
1. A model predicting who gets assigned to treatment.
2. A model predicting who belongs to which "club."
3. A model predicting the outcome.
The Magic: Their new method is "triply robust." This means you only need two of those three models to be correct to get the right answer. If you get two right, the third one can be completely wrong, and the math still works. It's like a bridge that stays standing even if one of its three support pillars collapses.

4. The "Machine Learning" Upgrade

The paper also shows how to use Machine Learning (AI) to build these models.

The Problem: Machine learning is great at finding patterns, but it can be "noisy" and introduce bias if you aren't careful.
The Solution: They use a technique called Cross-Fitting. Imagine you have a deck of cards. You split the deck into piles. You use one pile to teach the AI how to predict, and a different pile to test the prediction. You swap the piles around many times. This ensures the AI doesn't just "cheat" by memorizing the data, giving you a clean, unbiased result.

5. Real-World Test: The Job Corps Study

The authors tested this on real data from the Job Corps (a program for disadvantaged youth).

Scenario A (Employment): They looked at whether people got a job. They found that for people who would get a job regardless of the program, the program increased their chances of earning more money.
Scenario B (Ranking): They looked at income levels (Unemployed, Low Pay, High Pay). They found the program helped people "win" against the control group, pushing them into higher income brackets.
The Surprise: They found that even people who didn't actually join the program (but were assigned to it) seemed to do slightly better, perhaps because the offer of the program gave them a psychological boost.

Summary

This paper provides a robust, flexible toolkit for measuring cause-and-effect when:

People don't follow instructions perfectly.
The data is messy, skewed, or ranked (not just numbers).
We aren't 100% sure our statistical models are perfect.

It replaces the old "average difference" ruler with a "winning probability" scorecard, backed by a safety net that works even if some of our assumptions are slightly off.

Technical Summary: Principal Stratification with U-statistics under Principal Ignorability

Problem Statement
Principal stratification is a standard framework for causal inference when an intermediate variable (e.g., treatment noncompliance or survival status) exists. While the literature predominantly focuses on the principal average causal effect (PACE)—defined as the expected difference in potential outcomes within a specific principal stratum—this metric may be insufficient when the scientific interest lies in the relative ordering of outcomes rather than their mean differences. This is particularly relevant for ordinal outcomes, composite endpoints, or distributions with heavy tails and outliers where mean contrasts are sensitive. Furthermore, existing methods for generalized causal effects (such as probabilistic indices or win ratios) have not been fully extended to the principal stratification setting, leaving a gap in nonparametric identification and efficient estimation for these estimands.

Methodology
The authors propose a framework for estimating the Principal Generalized Causal Effect (PGCE), defined as $\tau^s_h = E[h\{Y_1(1), Y_2(0)\} | S_1 = S_2 = s]$ , where $h$ is an arbitrary contrast function (potentially nonlinear) and $S$ denotes the principal stratum. The methodology proceeds through the following steps:

Identification: Under three structural assumptions—Treatment Ignorability, Monotonicity (excluding the "defier" or "harmed" stratum), and Pairwise Mean Principal Ignorability—the authors derive nonparametric identification formulas. The Pairwise Mean Principal Ignorability assumption generalizes the unitwise assumption found in prior work (e.g., Jiang et al., 2022) to accommodate nonlinear contrast functions. It posits that, conditional on covariates, the expected contrast function is exchangeable across specific pairs of latent principal strata that map to the same observed data strata.
Efficient Influence Functions (EIFs): The authors derive the EIFs for the PGCE estimands. These functions characterize the semiparametric efficiency bound and serve as the basis for constructing optimal estimators. The derivation accounts for the ratio structure of many PGCEs (e.g., win ratios) and the dependence on nuisance functions defined on pairs of observations.
Estimation via U-statistics: Because the EIFs involve nuisance functions defined on pairs of observations, the authors construct estimators using U-statistics.
- Multiply Robust Estimators: They develop estimators that are "triply robust," meaning they remain consistent and asymptotically normal if any two of the three nuisance models (treatment propensity score, principal score, or pairwise outcome mean) are correctly specified.
- Debiased Machine Learning (DML): To achieve full asymptotic efficiency and accommodate flexible, data-adaptive nuisance function estimation, the authors integrate the U-statistic framework with cross-fitting (sample splitting). This allows for the use of machine learning algorithms (e.g., random forests, SuperLearner) while maintaining $n^{1/2}$ -consistency and asymptotic normality under mild regularity conditions.
Sensitivity Analysis: The paper addresses the untestable nature of the monotonicity assumption by proposing a sensitivity analysis strategy. This involves introducing a sensitivity function $\eta(X)$ to quantify the proportion of the "harmed" stratum, allowing for bias-corrected estimation when monotonicity is violated.

Key Contributions

Generalization of Estimands: The paper introduces the PGCE, a class of estimands that accommodates arbitrary contrast functions within principal strata, extending beyond the standard PACE to include probabilistic indices and win ratios.
Theoretical Extension: It expands the theoretical results of Jiang et al. (2022) from linear contrasts to a broader class of nonlinear contrasts under principal ignorability, providing the first nonparametric identification formulas for these generalized effects in the presence of a binary intermediate variable.
Efficient Influence Functions: The derivation of EIFs for PGCEs establishes the semiparametric efficiency lower bounds for these estimands.
Triply Robust U-statistics: The authors demonstrate that the proposed U-statistic-based estimators possess a triple robustness property (consistent if two of three models are correct) and achieve local efficiency when all models are correct.
Integration with Machine Learning: The work provides a rigorous theoretical foundation for applying debiased machine learning to principal stratification with nonlinear contrasts, proving that cross-fitted U-statistics achieve the efficiency bound under standard machine learning convergence rates.

Results
The proposed methods were evaluated through simulation studies and a real-data application:

Simulations: In scenarios with continuous and ordinal outcomes, the triply robust and debiased machine learning (DML) estimators demonstrated consistency even when one of the three nuisance models was misspecified. In contrast, standard plug-in moment estimators were biased under such misspecification. The DML estimators generally exhibited slightly lower bias and improved accuracy compared to parametric triply robust estimators, particularly in finite samples.
Data Example (Jobs Corps): The methods were applied to the U.S. Jobs Corps study to estimate the causal effect of job training on earnings, accounting for noncompliance and employment status (truncation by death).
- For the Survivor Probabilistic Index (SPI) among always-employed participants, the DML and triply robust methods estimated a probability of 0.539 that a treated individual has higher earnings than a control individual (95% CI excluding 0.5), suggesting a positive treatment effect.
- For the Principal Win Ratio (PWR) with ordinal earnings, the analysis revealed heterogeneous effects across strata. Compliers showed a significant positive effect (win ratio ~1.27), while never-takers showed a neutral effect. Notably, always-takers also showed a significant positive effect, suggesting a potential psychological impact of the assignment itself.
- Sensitivity Analysis: The PWR estimators showed strong robustness to violations of the monotonicity assumption, whereas the SPI estimates showed a slight shift in magnitude as the proportion of "harmed" strata increased.

Significance and Claims
The paper claims to provide a comprehensive framework for causal inference in principal stratification that moves beyond mean contrasts to accommodate ordinal and composite outcomes. By establishing the efficient influence functions and developing triply robust U-statistic estimators, the work fills a critical gap in the literature regarding the identification and efficient estimation of generalized causal effects. The authors emphasize that their approach offers a principled way to handle nonlinear contrasts and composite endpoints while maintaining robustness against model misspecification. The integration of machine learning via cross-fitting allows for flexible modeling of complex nuisance functions, making the framework applicable to modern high-dimensional data settings. The authors modestly note that the current development assumes independent and identically distributed (i.i.d.) data and that extending these methods to clustered or longitudinal data structures remains a direction for future research.