Principal Stratification in Causal Inference

• Principal stratification is a causal inference framework that defines subpopulations based on potential intermediate variables to reveal heterogeneous treatment effects.
• It relies on key assumptions like treatment ignorability, monotonicity, and uses copula models to handle cross-world dependencies.
• Recent advances include semiparametric estimation, doubly robust methods, and integration of machine learning techniques for continuous outcome settings.

Principal stratification is a foundational approach in modern causal inference for addressing post-treatment or intermediate variables that mediate or confound the relationship between treatment and outcome. By defining causal effects within subpopulations characterized by their joint potential values of these intermediates—so-called "principal strata"—the framework enables rigorous estimation and interpretation of causal effects in settings such as noncompliance, truncation by death, mediation, and surrogate endpoint evaluation. The methodology, originally introduced in the context of binary intermediate variables, has been significantly extended to accommodate increasingly complex scenarios, including those with continuous post-treatment variables, multiple or continuous treatments, and high-dimensional covariates.

1. Fundamental Concepts of Principal Stratification

The principal stratification framework operates in the potential outcomes paradigm. For a study unit with binary treatment $Z\in\{0,1\}$, continuous post-treatment variable $S \in \mathbb{R}$, and outcome $Y$, the standard notation is:

• $S_1, S_0$: potential values of $S$ under treatment and control,
• $Y_1, Y_0$: potential outcomes under treatment and control,
• Principal stratum $U = (S_1, S_0)$.

When $S$ is binary, $U$ has four possible values (e.g., always-taker, never-taker, complier, defier). With continuous $S$, $S \in \mathbb{R}$0 ranges over $S \in \mathbb{R}$1, leading to infinitely many principal strata.

The primary estimand is the principal-stratified causal effect surface:

$S \in \mathbb{R}$2

or, equivalently, the stratum-specific means $S \in \mathbb{R}$3 for $S \in \mathbb{R}$4 (Lu et al., 2023).

Such estimands decompose population-level or intention-to-treat effects into effects for scientifically relevant subpopulations—those classified by their joint potential intermediate values—thereby illuminating heterogeneity and underlying mechanisms.

2. Identification: Assumptions and Structural Functions

Identification of principal causal effects within latent principal strata is challenging because strata membership for any unit is fundamentally unobservable. Standard identification relies on the following assumptions (Lu et al., 2023, Jiang et al., 2020, Lipkovich et al., 2021):

• Treatment ignorability: $S \in \mathbb{R}$5. Assignment is randomized (or randomized given covariates).
• Principal ignorability: $S \in \mathbb{R}$6 and analogously for $S \in \mathbb{R}$7.
• Monotonicity (often): $S \in \mathbb{R}$8 for all units, ruling out "defiers."
• No interference (SUTVA): One unit's outcome does not depend on others’ treatment or intermediates.

For continuous $S \in \mathbb{R}$9, joint distributions $Y$0 and $Y$1 are identified, but their joint law is not. The dependence between $Y$2 and $Y$3 (the so-called "cross-world dependence") is modeled by a copula $Y$4, with density $Y$5 (Lu et al., 2023, Yang et al., 2024).

The principal-stratum density conditional on $Y$6 is:

$Y$7

where $Y$8.

Identification of the mean outcomes within stratum $Y$9, under these assumptions, is given by:

$S_1, S_0$0

with $S_1, S_0$1 (Lu et al., 2023).

Alternative identification strategies leverage auxiliary variables $S_1, S_0$2 under principal ignorability or auxiliary-independence conditions; nonparametric and semiparametric identification can be obtained using invertibility conditions, completeness arguments, and/or copula models (Jiang et al., 2020).

3. Estimation and Semiparametric Efficiency for Continuous Principal Strata

The curse of dimensionality and non-regularity make direct pointwise estimation of $S_1, S_0$3 infeasible in the continuous setting. Modern methodology addresses this by projecting the surface onto a finite-dimensional, smooth "working model" $S_1, S_0$4 with parameter $S_1, S_0$5:

$S_1, S_0$6

for some analytic weight $S_1, S_0$7 (Lu et al., 2023).

By differentiating the above functional and translating the identification formulas, one obtains a Z-estimation equation

$S_1, S_0$8

where $S_1, S_0$9 is the efficient influence function (EIF) corresponding to the model parameter $S$0. The EIF encompasses:

• Terms corresponding to perturbation of the marginal $S$1 distribution,
• Perturbation of the $S$2 densities (via principal score/copula structure),
• Perturbation of the $S$3 regression surface.

By plugging in estimated nuisance functions $S$4 and solving the EIF equation, one obtains an estimator that is:

• Doubly robust: consistent if the principal-score model is correct and at least one of the outcome or treatment models is correctly specified.
• Semiparametrically efficient: achieves the efficiency bound when all nuisance estimators converge sufficiently fast (i.e., root-$S$5 rates for parametric models or $S$6 for density estimation), with

$S$7

(Lu et al., 2023).

Smoothing or "localization" (kernel or working model) approaches ameliorate the inherent non-regularity of estimating a pointwise surface with infinite strata (Zhang et al., 2024). Localization ensures statistical regularity and enables the construction of estimators with minimax optimality, attaining $S$8 rates for sufficiently smooth targets.

4. Computational Implementation and Practical Considerations

The framework described above has been operationalized in software. The R package PSContinuous implements:

• Estimation routines for inverse-probability and outcome modeling (pd_om), treatment-probability weighting (tp_pd), and EIF-based estimation (eif_est).
• Estimation of the copula density, $S$9 (kernel or parametric), and regression functions $Y_1, Y_0$0 (linear models, generalized additive models).
• Bootstrap and EIF-based variance estimation and confidence intervals.

Recommendations include:

• Sensitivity analysis with respect to the copula parameter $Y_1, Y_0$1 is essential, as stratum-density identification is up to the copula.
• If estimated propensity scores $Y_1, Y_0$2 are near 0 or 1, certain weighting-based estimators become numerically unstable; outcome-model projection is often preferable with well-specified regressions.
• Kernel estimation for $Y_1, Y_0$3 is practical in low to moderate dimensions; higher-dimensional settings may require parametric or semiparametric density estimation (Lu et al., 2023).

5. Extensions: Bayesian Nonparametrics and Machine Learning

Alternative strategies for continuous post-treatment strata involve Bayesian nonparametric models—e.g., Dirichlet process mixtures and confounder-aware shared-atom mixtures—which allow data-adaptive clustering and smoothing of the infinite-dimensional causal effect surface (Zorzetto et al., 2024, Zorzetto et al., 2024). These methods enable a data-driven coarsening of strata, model-based regularization, and full uncertainty quantification of both strata membership and causal effects.

Further, machine learning methods for principal stratification—particularly doubly robust and debiased machine learning with cross-fitting and influence-function-based pseudo-outcomes—deliver root-$Y_1, Y_0$4 consistency and honest inference under data-adaptive nuisance estimates, extending beyond classical parametric approaches (Chen et al., 2024, Tong et al., 27 Jun 2026). Such methodologies facilitate discovery of heterogeneous principal causal effects in high-dimensional or functional covariate spaces.

6. Applications and Sensitivity Analyses

Principal stratification has been applied in diverse contexts, including:

• Surrogate endpoint evaluation, where the continuous surrogate (e.g., biomarker) is evaluated for causal necessity and sufficiency (Jiang et al., 2015, Lu et al., 2023).
• Mediation and mechanism assessment with time-varying intermediates in stepped-wedge cluster randomized trials, utilizing copula-based sensitivity analyses (Yang et al., 2024).
• Empirical studies in education, clinical trials (e.g., supportive home care on readmission and death (Comment et al., 2019)), and environmental regulation (Zorzetto et al., 2024, Zorzetto et al., 2024).

Sensitivity analysis with respect to copula parameters, cross-world dependence, and assumptions such as monotonicity is foundational. Analysts are advised to vary copula/link parameters and present causal effect estimates as functions of such sensitivity parameters (Lu et al., 2023, Zhang et al., 2024).

7. Summary

Principal stratification enables rigorous causal inference in complex settings involving post-treatment intermediates. For continuous post-treatment variables:

• Identification is achieved via principal ignorability plus a copula model on the bivariate distribution of potential intermediates.
• Semiparametric projection and EIF-based estimators attain double robustness and efficiency for projected mean surfaces or localized functionals.
• Bayesian and machine-learning approaches extend estimation and regularization to the high-dimensional, infinite-strata regime.
• Sensitivity analysis to cross-world dependence and model misspecification is essential in practice.

Ongoing research further generalizes principal stratification to continuous treatments, functional intermediates, and highly flexible heterogeneous effect surfaces, ensuring the approach remains at the frontier of causal methodology (Antonelli et al., 2023, Tong et al., 27 Jun 2026).