Principal Stratification

Principal stratification is a counterfactual causal inference framework in biostatistics, epidemiology, and public health introduced by Greek biostatistician Constantine E. Frangakis and American statistician Donald B. Rubin in 2002. Working within the potential outcomes (Rubin Causal Model) paradigm, principal stratification solves the problem of how to define and estimate valid causal effects when an investigator wishes to adjust for or restrict attention to a post-treatment intermediate variable S—such as treatment compliance, survival to a follow-up biomarker assessment, or acquisition of infection in a preventive vaccine trial—that occurs after treatment assignment Z and before (or concurrently with) the primary clinical outcome Y.

In a randomized controlled trial or observational cohort study, naively conditioning on the observed post-treatment variable S^obs (for instance, comparing outcomes Y among vaccinated individuals who became infected S^obs = 1 versus placebo recipients who became infected S^obs = 1) destroys baseline randomization and induces post-treatment selection bias—a form of collider bias in which treatment Z and unmeasured susceptibility U both affect S. Because the subset of individuals with S(1) = 1 under active treatment is a biologically different subpopulation from the subset with S(0) = 1 under control, contrasting E[Y | Z = 1, S^obs = 1] − E[Y | Z = 0, S^obs = 1] compares two non-identical sets of people and does not measure a causal effect on any fixed population. Frangakis and Rubin resolved this pathology by stratifying subjects not on their single observed post-treatment value S^obs, but on their joint counterfactual pair G = (S(0), S(1))—their principal stratum. Because an individual's potential intermediate responses (S(0), S(1)) under both fixed regimes are unaffected by which treatment Z ∈ {0, 1} the individual actually receives, principal strata behave mathematically as latent pre-treatment baseline covariates, ensuring that any comparison of Y(1) versus Y(0) within a union of principal strata is a genuine causal effect.

Why "Net Treatment Comparisons" Fail to Be Causal Effects

Consider a randomized clinical trial assigning i = 1, …, n participants to binary treatment Z ∈ {0, 1} (where Z = 1 denotes active intervention and Z = 0 denotes control). Under the Stable Unit Treatment Value Assumption (SUTVA: no interference between units and a single version of each treatment), each participant possesses two potential values of the post-treatment intermediate variable S(0) and S(1), alongside two potential values of the primary outcome Y(0) and Y(1). Observed variables are linked to potential outcomes by consistency:

Prior to Frangakis and Rubin (2002), clinical investigators frequently reported what Paul R. Rosenbaum (1984) termed net treatment comparisons (or post-infection / survivor subgroup contrasts), conditioning on the observed post-treatment realization S^obs = s:

Rewriting Δ_net(s) in terms of potential outcomes under physical randomization Z ⊥ (S(0), S(1), Y(0), Y(1)) exposes the flaw:

Whenever treatment Z has a non-null causal effect on the intermediate variable S (so that the subpopulation with S(1) = s differs from the subpopulation with S(0) = s), the two conditional expectations in Δ_net(s) average over two different sets of individuals:

1. In Preventive Vaccine Trials (Post-Infection Viral Load or Disease Severity): Let Z ∈ {0, 1} denote randomization to an HIV, malaria, or SARS-CoV-2 vaccine versus placebo, S ∈ {0, 1} denote breakthrough infection during follow-up, and Y denote post-infection viral load or hospitalization. If the vaccine partially protects against infection (P(S(1) = 1) < P(S(0) = 1)), participants who still acquire infection despite vaccination (S(1) = 1) tend to have weaker baseline mucosal immunity or higher exposure risk U than the broader pool who acquire infection under placebo (S(0) = 1). Conditioning on observed infection S^obs = 1 opens the collider path Z → S^obs ← U → Y (collider bias), making an effective vaccine appear to worsen post-infection viral load Y even when Y(1) = Y(0) for every person.
2. In Critical Care and Oncology Trials ("Truncation by Death"): Let S ∈ {0, 1} indicate survival to 12 months and Y denote health-related quality of life (QoL) at 12 months (which is undefined when S = 0). If a life-saving therapy Z = 1 rescues critically ill patients who would have died under control (S(0) = 0, S(1) = 1), those rescued patients enter the observed 12-month survivor pool with low baseline quality of life, pulling down the average QoL among treated survivors E[Y(1) | S(1) = 1] below E[Y(0) | S(0) = 1] even when treatment improves QoL for every patient who would survive under both arms.

The Four Canonical Binary Principal Strata

When both treatment assignment Z ∈ {0, 1} and the post-treatment intermediate variable S(z) ∈ {0, 1} are binary, the basic principal stratification partitions the study population into four mutually exclusive and exhaustive latent subpopulations defined by the ordered pair G = (S(0), S(1)) ∈ {(0, 0), (0, 1), (1, 0), (1, 1)}:

More generally, when S(z) is continuous (such as a post-treatment antibody titer), the basic principal stratification classifies units by the continuous ordered pair G = (S(0), S(1)) ∈ ℝ².

Invariance of Principal Strata to Treatment Assignment (Property 1 & Property 2)

Frangakis and Rubin (2002) established two foundational properties of any principal stratification:

1. Property 1 (Baseline Covariate Invariance): The principal stratum membership G = (S(0), S(1)) is unaffected by treatment assignment Z. If Z is randomly assigned (so Z ⊥ (S(0), S(1), Y(0), Y(1))), then randomization holds within every principal stratum:

1. Property 2 (Causal Validity of Principal Effects): A Principal Causal Effect (PCE) is defined as a comparison of potential outcomes Y(1) and Y(0) within a fixed principal stratum G = (s₀, s₁):

Because the conditioning event (S(0) = s₀, S(1) = s₁) defines the identical subpopulation under both Y(1) and Y(0), every Principal Causal Effect is a well-defined, collider-bias-free causal effect.

1. Local Average Treatment Effect (LATE / CACE) Under Non-Compliance

Prior to the general 2002 formulation, Joshua D. Angrist, Guido W. Imbens, and Donald B. Rubin (1996) developed the Complier Average Causal Effect (CACE)—known in econometrics and Mendelian randomization as the Local Average Treatment Effect (LATE)—which is the special case of principal stratification for binary treatment non-compliance (S(z) ∈ {0, 1} denoting actual treatment receipt).

In a trial with non-compliance, the target estimand is the causal effect of treatment among Compliers G = (0, 1): PCE(0, 1) = E[Y(1) − Y(0) | S(0) = 0, S(1) = 1]. Because we never observe both S(0) and S(1) for the same individual, each observed cell (Z = z, S^obs = s) mixes two latent principal strata:

• Observed (Z = 1, S^obs = 1) contains Always-Takers (1, 1) and Compliers (0, 1).
• Observed (Z = 0, S^obs = 0) contains Never-Takers (0, 0) and Compliers (0, 1).
• Observed (Z = 1, S^obs = 0) contains Never-Takers (0, 0) and Defiers (1, 0).
• Observed (Z = 0, S^obs = 1) contains Always-Takers (1, 1) and Defiers (1, 0).

Angrist, Imbens, and Rubin achieved point identification of CACE via two structural assumptions:

1. Monotonicity (No Defiers): S(1) ≥ S(0) for all individuals, so π₁₀ = P(G = (1, 0)) = 0. Under monotonicity, observed (Z = 1, S^obs = 0) consists only of Never-Takers (0, 0) (identifying π₀₀ = P(S^obs = 0 | Z = 1)), observed (Z = 0, S^obs = 1) consists only of Always-Takers (1, 1) (identifying π₁₁ = P(S^obs = 1 | Z = 0)), and the Complier proportion is π₀₁ = E[S^obs | Z = 1] − E[S^obs | Z = 0].
2. Exclusion Restriction for Always-Takers and Never-Takers: Randomized assignment Z has zero direct effect on Y within strata where treatment receipt does not change: PCE(0, 0) = PCE(1, 1) = 0.

Decomposing the intention-to-treat effect E[Y^obs | Z = 1] − E[Y^obs | Z = 0] = ∑ π_s0s1 PCE(s₀, s₁) yields the instrumental variable Wald ratio:

2. Survivor Average Causal Effect (SACE) Under Truncation by Death

When the primary outcome Y (such as cognitive function or quality of life) is only defined for surviving patients (S(z) = 1) and is truncated by death (Y(z) = ∗ when S(z) = 0)—a distinct problem from competing risks analysis of event times—only in the Always-Survivors stratum G = (1, 1) are both potential outcomes Y(0) and Y(1) simultaneously well-defined real numbers. Donald B. Rubin (2006) therefore established the Survivor Average Causal Effect (SACE) as the unique well-defined causal contrast on quality of life:

3. Vaccine Post-Infection Effects and Principal Surrogates

Similarly, in preventive vaccine trials (Hudgens, Hoering, and Self 2003), post-infection viral load or progression to severe disease Y is defined only among infected participants (S(z) = 1), so the causal effect of vaccination on post-infection severity is defined within the Always-Infected stratum G = (1, 1): PCE_inf = E[Y(1) − Y(0) | S(0) = 1, S(1) = 1].

For biomarker surrogate validation, Frangakis and Rubin (2002) replaced Ross L. Prentice's post-treatment statistical surrogate criteria with principal surrogacy: a biomarker S(z) is a valid principal surrogate for clinical outcome Y if PCE(s, s) = E[Y(1) − Y(0) | S(0) = s, S(1) = s] = 0 for all s (causal necessity) and PCE(s₀, s₁) increases across strata with larger biomarker shifts s₁ − s₀.

Mixture Deconvolution Under Monotonicity

In SACE and vaccine post-infection analyses, the exclusion restriction does not hold (treatment is expected to affect Y inside G = (1, 1)). Under survival monotonicity (S(1) ≥ S(0), so π₁₀ = 0):

1. Every control survivor (Z = 0, S^obs = 1) belongs to the Always-Survivors stratum (1, 1), directly identifying the control mean E[Y(0) | G = (1, 1)] = E[Y^obs | Z = 0, S^obs = 1] and stratum proportion π₁₁ = P(S^obs = 1 | Z = 0).
2. Treated survivors (Z = 1, S^obs = 1, with probability p₁ = π₁₁ + π₀₁) are a two-component mixture of Always-Survivors (1, 1) (fraction q = π₁₁ / p₁) and Protected Survivors (0, 1) (fraction 1 − q = π₀₁ / p₁):

Three Methods to Resolve the Mixture

1. Sharp Non-Parametric Bounds (Hudgens, Hoering, and Self 2003; Zhang and Rubin 2003): The sharp lower bound on E[Y(1) | G = (1, 1)] attributes the lowest q = π₁₁ / p₁ quantile tail of the observed Y^obs distribution among treated survivors to Always-Survivors (1, 1), and the sharp upper bound attributes the highest q quantile tail to Always-Survivors.
2. Sensitivity Parameter Tilting (Gilbert, Bosch, and Hudgens 2003): Specify a logistic selection parameter β indexing how Y(1) differs between Always-Survivors (1, 1) and Protected Survivors (0, 1) via P(G = (1, 1) | S(1) = 1, Y(1) = y) = [1 + exp(−α − βy)]⁻¹ and trace SACE(β) across plausible values of β.
3. Principal Ignorability and Principal Scores (Ding and Lu 2017): When rich pre-treatment baseline predictors X of survival S are measured, principal ignorability assumes Y(1) ⊥ S(0) | (S(1) = 1, X), identifying SACE via principal score weighting w(X) = P(S = 1 | Z = 0, X) / P(S = 1 | Z = 1, X).

How does principal stratification differ from causal mediation analysis (natural direct and indirect effects)?

Causal mediation analysis defines the natural direct effect E[Y(1, S(0)) − Y(0, S(0))] by physically intervening to set the intermediate variable S to its control counterfactual value S(0). In truncation by death (S = survival) or vaccine breakthrough infection (S = infection), it is meaningless to "intervene to set survival S to 1" after death, and Y(1, S(0)) relies on untestable cross-world independence assumptions. Principal stratification never intervenes on S; it evaluates Y(1) − Y(0) within latent baseline subgroups defined by (S(0), S(1)).

Can we ever know which specific individual belongs to the Always-Survivors stratum G = (1, 1)?

Because we observe S(1) for treated participants and S(0) for control participants, individual membership is only partially observed (under monotonicity, control survivors with S(0) = 1 are known to be Always-Survivors, and treated non-survivors with S(1) = 0 are known to be Never-Survivors). Even though individual stratum membership is latent among treated survivors, the population-level causal effect E[Y(1) − Y(0) | G = (1, 1)] is bounded or identified from marginal distributions and baseline covariates X.

Why is SACE defined only on the Always-Survivors stratum G = (1, 1)?

As Rubin (2006) demonstrated, restricting quality-of-life comparisons to G = (1, 1) reflects the mathematical reality of truncation by death: for any patient outside G = (1, 1), at least one of the two potential quality-of-life outcomes Y(0) or Y(1) does not exist because the patient is deceased under that arm.

How does Mendelian randomization relate to principal stratification?

When a binary genetic variant Z ∈ {0, 1} serves as an instrument for a binary exposure S ∈ {0, 1} in Mendelian randomization under monotonicity, the Wald ratio identifies the Principal Causal Effect inside the genetic Complier stratum G = (S(0) = 0, S(1) = 1).

1. Frangakis, C. E., & Rubin, D. B. (2002). "Principal stratification in causal inference." Biometrics, 58(1), 21–29.
2. Angrist, J. D., Imbens, G. W., & Rubin, D. B. (1996). "Identification of causal effects using instrumental variables." Journal of the American Statistical Association, 91(434), 444–455.
3. Hudgens, M. G., Hoering, A., & Self, S. G. (2003). "On the analysis of viral load endpoints in HIV vaccine trials." Statistics in Medicine, 22(14), 2281–2298.
4. Rubin, D. B. (2006). "Causal inference through potential outcomes and principal stratification: Application to studies with 'censoring' due to death." Statistical Science, 21(3), 299–309.
5. Ding, P., & Lu, J. (2017). "Principal stratification analysis using principal scores." Journal of the Royal Statistical Society: Series B (Statistical Methodology), 79(3), 757–777.