📊 statistics

Beyond principal ignorability: Nonparametric sensitivity bounds for principal stratification

This paper introduces a nonparametric sensitivity analysis framework for principal stratification that derives sharp bounds for principal causal effects under violations of the untestable principal ignorability assumption, utilizing a margin-free bounding factor and extending the methodology to generalized causal effects and pairwise comparisons.

Original authors: Xinyuan Chen, Michael O. Harhay, Fan Li

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinyuan Chen, Michael O. Harhay, 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 a detective trying to figure out if a new medicine (Treatment) actually cures a disease (Outcome). But there's a twist: before you see if the patient gets better, they have to take a pill that might make them feel sick (Intermediate Variable).

In the world of statistics, this is called Principal Stratification. It's a way of grouping people based on how they would react to the pill, not just how they actually did. For example:

The "Always-Takers": People who would take the pill no matter what.
The "Never-Takers": People who would never take it.
The "Compliers": People who only take it if you tell them to.
The "Defiers": People who do the opposite of what they're told.

The problem is, you can't see these groups directly. You only see who actually took the pill and who got better. To figure out the true effect of the medicine on the "Compliers," statisticians usually have to make a big guess called Principal Ignorability. This guess says: "The reason someone is a 'Complier' has nothing to do with how sick they are or how well the medicine works for them."

The Problem: This guess is untestable. It's like assuming a suspect is innocent just because you haven't found a motive yet. If this guess is wrong, your conclusion about the medicine could be completely wrong.

The Solution: A "Sensitivity Safety Net"

This paper introduces a new tool to check how strong that guess needs to be before your conclusion falls apart. Think of it as a stress test for your detective work.

Instead of just saying "We assume the guess is true," the authors ask: "How much would the hidden motive (unmeasured confounder) have to lie to make our conclusion disappear?"

They use two main "levers" to measure this hidden motive:

The Selection Lever: How much does the hidden motive change the type of person someone is? (e.g., Does being sick make you more likely to be a "Complier"?)
The Outcome Lever: How much does the hidden motive change the result? (e.g., Does being sick make the medicine work better or worse?)

They combine these into a single number called a Bounding Factor.

If this number is 1, it means there is no hidden motive, and your original guess is perfect.
If the number is high, it means a very strong hidden motive would be needed to ruin your conclusion.
If the number is low, it means even a tiny hidden motive could destroy your finding.

The "Nested" Discovery

The authors found something really cool about their new tool. Imagine a set of Russian nesting dolls:

The outermost doll is the "Worst-Case Scenario." This assumes the hidden motive is as evil as possible, making the truth as unclear as it can be.
The innermost doll is their new, precise calculation.

They proved that their new, precise bounds always fit inside the worst-case bounds. As the "evilness" of the hidden motive increases, their precise bounds slowly expand until they touch the walls of the worst-case scenario. This connects their new method to older, more conservative methods, showing they are all part of the same family.

The "E-Value" (The "How Strong?" Meter)

To make this easy for people to understand, they created a metric called the Principal E-value.

Think of this as a "strength rating" for the lie you'd have to tell to make your result vanish.
If your E-value is 2, it means an unmeasured factor would have to be twice as strong as the strongest factor you already measured to cancel out your result.
If your E-value is 1.1, your result is very fragile; a tiny, almost invisible factor could ruin it.

Two Real-World Tests

The authors tested their tool on two real datasets:

Housing and Health: They looked at whether living in damp houses causes depression. They found that for people who only get sick when living in damp conditions, the result was quite robust. The "hidden motive" would have to be incredibly strong to make the result disappear.
Job Training: They looked at whether a job training program helped people earn more money. Here, the results were much more fragile. For some groups, even a moderate hidden factor (like age or education) could explain away the results. This tells researchers: "Be careful! Your conclusion here is shaky."

Going Deeper: The "Pairwise" Challenge

The paper also tackles a harder version of the problem. Instead of comparing "Before vs. After" for one person, imagine comparing two different people to see who did better. This is useful for things like "Who is more likely to survive?" or "Who has a better chance of winning?"

The authors showed that when you compare two people, the math gets trickier because the "hidden motive" affects both people at once. They developed a new set of rules for this "pairwise" comparison, proving that these comparisons are even more sensitive to hidden selection biases than single-person comparisons.

The Bottom Line

This paper doesn't tell you what the answer is; it tells you how much you can trust the answer. It gives researchers a way to say, "Our conclusion holds up unless there is a hidden factor that is stronger than X." It turns a vague assumption into a concrete, measurable safety check.

Technical Summary: Beyond Principal Ignorability: Nonparametric Sensitivity Bounds for Principal Stratification

Problem Statement
Principal stratification is a robust framework for estimating causal effects in the presence of intermediate variables (e.g., non-compliance, truncation by death) by partitioning the population into latent strata based on potential intermediate outcomes. However, point identification of Principal Causal Effects (PCEs) typically relies on the Principal Ignorability (PI) assumption, which posits that potential final outcomes are conditionally independent of latent stratum membership given observed covariates. This assumption is fundamentally untestable. Existing sensitivity analysis strategies often rely on Bayesian priors, specific distributional assumptions, or structural constraints like monotonicity and exclusion restrictions. There is a need for a nonparametric framework that quantifies violations of PI without these restrictive assumptions, providing sharp bounds and falsification thresholds for both standard PCEs and more complex estimands.

Methodology
The authors develop a nonparametric sensitivity analysis framework that relaxes the PI assumption by introducing an unmeasured confounder $U$ . The methodology is built upon three core components:

Sensitivity Parameters: The framework introduces two margin-free relative risk (RR) parameters to characterize the degree of PI violation:
- Selection RR ( $RR_{SU}$ ): The maximum imbalance in the prevalence of $U$ between a target principal stratum and a nuisance stratum within a covariate stratum.
- Outcome RR ( $RR_{UY}$ ): The maximum impact of $U$ on the expected potential outcome.
  These parameters generalize the confounder-based bounding factors of Ding and VanderWeele (2016a, 2016b) to the principal stratification setting.
Bounding Factor Derivation: The authors derive a cell-specific bounding factor, $BF_U$ , which is a deterministic function of $RR_{SU}$ and $RR_{UY}$ . Unlike standard settings, the target stratum mean is latent and enters the observed cell mean as a weighted mixture with a nuisance stratum. The mixing weights depend on the principal scores, which are further determined by the cross-world odds ratio parameter $\theta(x)$ (governing the dependence between $D(1)$ and $D(0)$ ). The derivation accounts for this mixture structure and the additional sensitivity dimension introduced by $\theta(x)$ .
Sharp Nonparametric Bounds:
- For PCEs: The authors derive sharp bounds for the conditional and marginal PCEs ( $\Delta_s$ ) by propagating the bounding factor through the mixture decomposition. These bounds are shown to nest within the "worst-case" nonparametric bounds of Grilli and Mealli (2008) and Long and Hudgens (2013), converging to them only as the bounding factor approaches infinity.
- For PGCEs: The framework is extended to Principal Generalized Causal Effects (PGCEs), defined by general nonlinear contrast functions $h(\cdot)$ over a product space (e.g., probabilistic indices). Due to the non-separability of nonlinear functions, the outcome RR must be defined jointly over unmeasured confounders of two independent replicates. This results in three distinct bounding factors ( $BF_{[1]}, BF_{[2]}, BF_{[1,2]}$ ) reflecting the multiplicative compounding of selection penalties.

Key Results and Theoretical Findings

Nesting Property: The proposed bounds strictly nest within existing worst-case bounds for any finite bounding factor. As the bounding factor diverges, the proposed bounds converge to the worst-case bounds, establishing a continuous algebraic bridge between point identification (under PI) and maximum adversarial confounding.
Cornfield-Type Conditions and E-values: The paper derives critical falsification thresholds (Cornfield-type conditions) required to nullify a target effect.
- For PCEs, it establishes a "Principal E-value," defined as the minimum joint magnitude of selection and outcome RRs required to explain away the effect.
- For PGCEs, the analysis reveals three distinct thresholds: a selection bias lower threshold, an outcome bias lower threshold, and a global maximum threshold. The ordering of these thresholds ( $\gamma^*_{SU} < \gamma^*_{UY} < \Gamma^*$ ) formalizes that pairwise estimands are structurally more vulnerable to selection confounding than to outcome confounding due to the multiplicative nature of independent selection penalties.
Monotonicity: Under the monotonicity assumption, the bounds simplify, and the relationship with existing worst-case bounds (Long and Hudgens, 2013) is preserved.
Binary Outcomes: The framework provides sharp bounds for multiplicative estimands (Risk Ratios and Odds Ratios) and demonstrates that the falsification thresholds are invariant across causal contrast scales (additive vs. multiplicative).

Empirical Illustrations
The methodology is applied to two real-world datasets:

WHO-LARES Study: Analyzing the effect of damp housing on depression. The authors quantify the robustness of the causal relative risk among the "harmed" stratum, showing that observed covariates (e.g., employment, smoking) do not reach the magnitude required to nullify the effect, suggesting the finding is robust to moderate unmeasured confounding.
U.S. Job Corps Study: Analyzing the effect of training on earnings (ordinal outcome) using a PGCE (probabilistic index). The analysis reveals that for the female subgroup, the estimated effect is highly sensitive; known covariates (age, education) exceed the required confounding magnitude to nullify the effect, indicating structural fragility in the causal identification for this group.

Significance and Claims
The paper claims to provide the first nonparametric sensitivity analysis framework for principal stratification that:

Operates without parametric assumptions, Bayesian priors, monotonicity, or exclusion restrictions.
Unifies existing disparate strands of literature (e.g., worst-case bounds vs. PI-based bounds) through a continuous sensitivity parameter.
Extends sensitivity analysis to the broader class of PGCEs, addressing a gap in handling complex outcomes (ordinal, categorical) and nonlinear contrasts.
Introduces novel theoretical insights regarding the structural vulnerability of pairwise estimands to selection confounding compared to outcome confounding.
Offers actionable tools (sharp bounds, E-values, falsification contours) for researchers to assess the robustness of causal claims in the presence of unmeasured confounding between latent strata and outcomes.

The authors maintain modesty regarding future directions, noting that extending the framework to continuous intermediate variables and incorporating debiased machine learning for high-dimensional covariate adjustment are necessary next steps not yet addressed in this work.