Causal Inference for All: Marginal Estimands for Outcomes Truncated by Death

arXiv:2607.00222 · stat.ME, math.ST, stat.TH · Submitted 2026-06-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Causal Inference for All".

Jane: Here's the gist of the paper. In longitudinal studies, outcomes are often truncated by death, meaning that they are only observed or well-defined conditional on intercurrent events such as survival <ref:2607.00222#pg2> The authors introduce methodology for a new set of estimands that (i) concern the entire population, (ii) remain causally interpretable, and (iii) leverage the longitudinal data commonly available in studies with outcomes truncated by death <ref:2607.00222#pg4>.

Tom: Defining New Estimands The paper introduces a set of estimands that characterize the potential outcome trajectories of each individual across two distinct survival periods: the guaranteed survival period and the extended survival period <ref:2607.00222#pg10>.

Jane: The guaranteed survival period, 0 ≤ t ≤ Ti(1) ∧ Ti(0), is defined as the period during which the individual would be alive under both treatments.

Tom: The extended survival period, Ti(1) ∧ T(0) < t ≤ Ti(1) ∨ T(0), is the additional survival time under the treatment that would prolong survival.

Paper summary: Tom: We started by looking at why these truncations are a problem. The paper points out that outcomes in longitudinal studies are often only well-defined conditional on survival, and this restriction creates a fundamental challenge for causal inference because contrasts restricted to those who survive don't necessarily have a causal interpretation.

Jane: So, the core thesis is that existing strategies have a trade-off: you can get causally interpretable estimands that target a specific subgroup of survivors, or you can use while-alive and composite summaries that apply to the whole population but are hard to interpret as true causal effects on non-mortality outcomes.

Lu: The authors address this by introducing methodology for a new set of estimands specifically designed to concern the entire population, remain causally interpretable, and leverage the longitudinal data we often have when outcomes are truncated by death.

Meng: They introduce these new estimands that characterize individual outcome trajectories across two distinct survival periods: the guaranteed survival period and an extended survival period under a treatment.

Lalam: The guaranteed survival period is defined as the time from zero up to Ti one and Ti zero, which is the time an individual would be alive under both treatments.

Tom: Then there's the extended survival period, which is defined as that additional survival time under the treatment that would prolong someone's life beyond what they'd survive otherwise.

Jane: They then classify these into three main groups: 'while guaranteed-survival' estimands, 'while extended-survival' estimands focusing on that extra survival time, and single-world marginal separable effects.

Lu: The while guaranteed-survival estimands summarize treatment effects over the period from zero up to Ti one and Ti zero, aiming to keep the full population appeal of while alive and composite summaries.

Meng: And then there are the while extended-survival estimands which focus specifically on that additional survival time under the treatment that extends life.

Lalam: Finally, they have single-world marginal separable effects which generalize conditional separable effects to cover the full population by comparing interventions that vary an outcome component while holding a survival component fixed.

Tom: Their identification and estimation results are developed using a substitution-variable approach for those first two classes, and an influence function-based estimator for the marginal separable ones under specific treatment decompositions.

Jane: For identification of those marginal separable effects, they achieve it by solving a system of four equations that come from the observed distribution.

Lu: And when they apply this methodology to a reanalysis of a prostate cancer trial, they found that the estimated benefit of one treatment compared to another depends on the clinical goal you are interested in.

Meng: That’s a key finding because it shows that DE appears more favorable when you emphasize exit-time or end-of-life outcomes, which is useful for setting realistic expectations.

Conclusion: Tom: So, to wrap up 'Causal Inference for All: Marginal Estimands for Outcomes Truncated by Death', the paper provides a new way to estimate causal effects in these tricky longitudinal settings by focusing on whole-population estimands that are interpretable.

Jane: The authors introduce these three classes of estimands—guaranteed survival, extended survival, and marginal separable effects—to bridge the gap between survivor-restricted estimates and full-population summaries.

Lu: It’s about making sure that even when you only have data while someone is alive, you can still get causal conclusions that apply to everyone.

Meng: The implication for deployment is that we might be able to use these tools to make decisions in clinical settings where survival truncation is a major factor, provided the specific assumptions about monotonicity hold true for the data.

Lalam: Ultimately, this work gives us concrete formulas—like those derived in Section four point three—for estimating these new estimands using maximum likelihood estimators, which are triply robust under certain constraints.

Tom: It’s a lot of technical machinery, but the paper shows that the method is sound for addressing the challenge of survival truncation by providing interpretable ways to look at longitudinal data.

Jane: So, it moves us toward a framework where we can handle these real-world longitudinal datasets without immediately having to restrict our conclusions only to those who happen to survive.

Ruixuan Zhao, Mats Stensrud, Linbo Wang

Department of Computer and Mathematical Sciences, University of Toronto Scarborough · Institute of Mathematics, Ecole Polytechnique Fédérale de Lausanne · Department of Statistical Sciences, University of Toronto

stat.ME, math.ST, stat.TH

Submitted: 2026-06-30

Updated: 2026-11-07

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: In longitudinal studies, outcomes are often truncated by death, meaning that they are only observed or well-defined conditional on intercurrent events such as survival <ref:2607.00222#pg2> The…

Key concepts

Guaranteed Survival Period
This is the time interval during which an individual is expected to be alive under both potential treatments. It defines the baseline survival window where both treatment paths are fully observable, making it useful for summarizing effects over that initial period.
Extended Survival Period
This refers to the additional survival time that one treatment might prolong compared to another. This estimand focuses specifically on the extra duration an individual survives under a particular intervention, rather than just the total survival time.
Marginal Separable Effects
This class of estimands compares interventions by varying an outcome-related component (ZY) while keeping a survival-related component (ZS) fixed. This generalizes previous work to the full population and allows researchers to isolate the effect of one factor independent of others.

Terminology

Summary

In longitudinal studies, outcomes are often truncated by death, meaning that they are only observed or well-defined conditional on intercurrent events such as survival <ref:2607. p. 000222#pg2> The authors introduce methodology for a new set of estimands that (i) concern the entire population, (ii) remain causally interpretable, and (iii) leverage the longitudinal data commonly available in studies with outcomes truncated by death <ref:2607. p. 000222#pg4>.

Defining New Estimands

The paper introduces a set of estimands that characterize the potential outcome trajectories of each individual across two distinct survival periods: the guaranteed survival period and the extended survival period <ref:2607.00222#pg10>. The guaranteed survival period, 0 ≤ t ≤ Ti(1) ∧ Ti(0), is defined as the period during which the individual would be alive under both treatments. p. 8 The extended survival period, Ti(1) ∧ T(0) < t ≤ Ti(1) ∨ T(0), is the additional survival time under the treatment that would prolong survival. p. 8

Classification of Estimands

The paper develops three main classes of estimands:

  1. “while guaranteed-survival” estimands, which summarize treatment effects over the guaranteed survival period [0, T(0) ∧ T(1)], retaining the full-population appeal of while-alive and composite summaries. p. 3

  2. “while extended-survival” estimands for the period during which an individual would survive under one treatment but not under the other, focusing on the additional survival time. p. 3

  3. single-world marginal separable effects that generalize recent work on separable effects, comparing interventions that vary an outcome-related component ZY while holding a survival-related component ZS fixed. p. 3

Identification and Estimation Results

The paper develops identification and estimation results for these estimands using a substitution-variable approach for the while guaranteed-survival and while extended-survival estimands. For marginal separable effects, they construct influence function-based estimators under specific treatment decompositions. p. 3 The identification of the components in (4) is achieved by solving a system of four equations derived from the observed distribution. p. 5

Empirical Application and Comparison

The methodology is applied in a reanalysis of a prostate cancer trial, highlighting how different estimands can yield different treatment conclusions. The results show that the estimated benefit of DE relative to MP depends on the clinical goal of interest, with DE appearing more favorable when emphasis is placed on exit-time or end-of-life outcomes. p. 9

Limitations and Future Directions

Several limitations remain, including the fact that while guaranteed-survival and while extended-survival estimands are full-population estimands but remain cross-world, and their identification relies on assumptions such as monotonicity, principal ignorability, and substitution-variable conditions. Future work may develop sensitivity analyses for these assumptions and extend the framework to continuous-time measurements, competing intercurrent events, and dynamic treatment regimes.

Simulation Validation

Simulation studies assess the numerical performance of the proposed estimators in Sections 4.3 and 5.4, reporting mean bias (scaled by 10 3) and standard error (scaled by 10 3). p. 0 The simulation results for the while guaranteed-survival estimands show that the exit-time estimate is close to the commonly used [-5, 5] benchmark for a clinically meaningful difference in QoL. p. 2 The conditional separable effect approach also provides estimates where the marginal separable contrast is larger than the corresponding estimate of µ ext. p. 3

Marginal Separable Effects

The paper introduces a new class of marginal separable effect estimands, Γ(zY, zS), which extends the CSE to the full population. This class allows for comparison between interventions that vary ZY while holding the survival-related component fixed at ZS = zS, over [0, T(zS)]. p. 3 The identification of these components is achieved under assumptions like Assumption 12 and 13 in the four-arm trial C. p. 7

Connection to Other Estimands

The paper establishes connections between the while guaranteed-survival and while extended-survival estimands and marginal separable effect estimands under certain assumptions. Specifically, it shows that ∆gua = ∆sep(0) under appropriate choices of weighting schemes. p. 3 Furthermore, when choosing appropriate weighting schemes to establish the equivalence, µ ext = Γ(1, 1) − Γ(1, 0). p. 1

Estimation Strategy

The estimation of these estimands involves constructing plug-in estimators using maximum likelihood estimators for the underlying model parameters. The resulting estimators are triply robust when solved under P4, P5, and P6 constraints. p. 1 The final plug-in quantities for µ(z) and µ ext are given by the formulas derived in Section 4. p. 33.

Influence Function Characterization

The nonparametric influence function of the estimand (17) is given by a complex expression involving several terms related to the latent processes. The derivation involves decomposing the influence function into four components, T1 through T4, which relate to different aspects of the potential outcomes and survival probabilities. p. 6

Simulation Details

The simulation studies utilize a specific data-generating mechanism involving latent processes TD, TH, and TP. The latent process formulation involves conditional probabilities of these three processes not occurring during interval r, given that they have not occurred before interval r.

Data Availability

Code for the simulation studies and real-data analysis is available at https://github.com/Ruixuan-Zhao/marginal-estimands-truncated-by-death. p. 0 The real data are not publicly available, so a simulated dataset is provided for running the real-data analysis code. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can calculate full-population causal estimands for outcomes truncated by death, such as while guaranteed-survival and while extended-survival estimands, which retain the full-population appeal of while-alive and composite summaries but define a coherent causal contrast.

  2. Improved AI systems can develop single-world marginal separable effect estimators that generalize recent work on separable effects to the full population by comparing interventions that vary an outcome component while holding a survival component fixed, using the formula Gamma(zY, zS).

  3. Improved AI systems can learn these proposed estimands from observed longitudinal trial data by applying identification results derived from substitution-variable approaches and constructing plug-in estimators based on observed data likelihood constraints.

  4. Improved AI systems can perform robust estimation under high-dimensional, continuous settings by using parameterizations like P1, P2, P3, and P4 to satisfy monotonicity and survival-ordering restrictions implied by the identification assumptions.

  5. Improved AI systems can utilize a triply robust estimator for these components by solving estimating equations derived from models defined in P5 and P6, or a simpler version omitting the term dependent on pLrSr−1,L¯r−1,Z,L0.

Sources

Related papers