Time-Varying Causal Survival Learning under Staggered Adoption
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: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Time-Varying Causal Survival Learning under Staggered Adoption".
Jane: Here's the gist of the paper. This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a fundamental challenge in medical research exemplified by the Stanford Heart Transplant study.
Tom: Problem Formulation and Framework The core problem addressed is estimating causal effects in settings where treatment timing varies across patients, such as in heart transplantation where waiting times are random.
Jane: Traditional causal inference fails because the treatment timing is random and highly variable, making it impossible to establish a fixed time point to separate control and treatment groups.
Tom: The authors extend staggered adoption designs by integrating survival analysis techniques, specifically using hazard functions to model the instantaneous probability of an event.
Title and authors: Tom: So we're looking at this paper today titled "Time-Varying Causal Survival Learning under Staggered Adoption." We're starting by talking about who wrote it and what the title really means.
Jane: It’s an interesting title because it immediately signals that they are dealing with treatment timing that changes over time, which is a major hurdle in causal inference.
Lu: The authors are Xiang Meng and Iavor Bojinov, and they are clearly bringing together expertise from both machine learning and causal modeling.
Meng: I'm interested in the authors because if this is going to be used in production, the methodology has to be solid enough to handle real-world messiness without needing constant retraining.
Tom: Exactly. So what does it mean for the listener? It means they are tackling that fundamental problem where treatment timing varies randomly, especially in complex medical research.
Jane: They're addressing that challenge by looking at settings like heart transplantation where waiting times are random because of things like donor availability or patient readiness.
Lu: The paper identifies specific conditions under which staggered adoption assumptions can justify the use survival analysis techniques for causal inference with time-varying treatments, which is a big theoretical move.
Tom: It means they are showing us that there’s a way to use established survival analysis tools while still keeping the causal interpretation clear, even with this time variation.
Jane: It’s about taking existing methods and making them more applicable to situations where the treatment isn't applied at a single fixed point.
The paper's summary: Tom: Now we get into what the paper actually says about the method, beyond just the setup. It summarizes how this Time-Varying Causal Survival Learning under Staggered Adoption works in plain terms.
Jane: Essentially, they are providing a formal way to estimate causal effects when treatment is applied at different times across different patients in a staggered fashion.
Lu: They use the framework of hazard functions to model the instantaneous probability of an event, which is key because survival data only gives us one outcome per patient.
Meng: So they take that hazard function and incorporate it with the treatment timing variable to get a dynamic estimate of how much the treatment affects survival over time.
Tom: The core idea is modeling it as at t times tau x plus eta zero x, where at t is the probability of adoption by time t, and eta zero x represents that baseline log hazard function.
Jane: This captures the dynamic nature of treatment assignment over time and allows us to see how the treatment effect evolves dynamically with patient characteristics.
Lu: They then proceed through a two-stage process where they first estimate those propensity scores, and then use those estimates to solve for the treatment effect function by minimizing a specific objective function.
Meng: This sounds like complex modeling that needs very good data quality just to avoid garbage results at scale, right?
Tom: It does sound complex, but they argue that if you have good data, this is how you get an unbiased and efficient estimation of the treatment effect function tau x by solving that minimization problem.
Jane: So for someone driving or cooking, this means we are building a system that can account for when the "treatment" actually happens along with who the patient is.
Lu: It's about making sure the model correctly links the patient's characteristics to their survival outcome over time in a causal way.
The paper's improvements: Tom: Let's talk about what they explicitly suggest as improvements to this approach, beyond just the core framework itself.
Jane: They point out a key limitation is that the current model assumes linear treatment effect heterogeneity, meaning it might be too simple for some real-world scenarios.
Lu: That’s where future work could extend the framework to use more flexible specifications for that treatment effect function tau x, maybe using non-parametric or semi-parametric approaches.
Meng: If they move to those more flexible specs, how does that impact the deployment cost? Does it make the model significantly harder to maintain in production?
Tom: It might increase complexity, but they also suggest considering two distinct hazard functions post-transplant to capture the immediate risk increase followed by a lower risk level.
Jane: That makes sense because in survival data, you often see different patterns—like an initial spike in risk right after an intervention followed by a stabilization.
Lu: That would allow the model to capture that dynamic pattern better than just one single hazard function for the whole duration.
Tom: They also suggest developing testing procedures for identifying which patients benefit most from transplantation, which moves beyond just estimating average effects.
Jane: So they're moving toward finding personalized insights, instead of just getting a general number on how effective the treatment is overall.
Conclusion: Tom: So to wrap up, this paper on Time-Varying Causal Survival Learning under Staggered Adoption really shows how robust this approach is when tested against simulations and real data from heart transplant studies.
Jane: It's a powerful tool because it provides a statistically rigorous way to analyze staggered treatment adoption with time-varying interventions in diverse applied contexts.
Lu: The paper proves that you can use survival analysis techniques for causal inference even when treatment timing is messy.
Meng: From my side, the practical implication is that we need models that can handle this kind of uncertainty without needing to constantly retrain on specific, curated datasets for every single application.
Lalam: I'm pretty sure the core finding is about building a robust estimator using Double Machine Learning to ensure unbiased and efficient estimation of treatment effects.
Tom: It really does give us reliable insights for optimizing customer acquisition and retention strategies across industries by accounting for both intervention timing and unit characteristics.
Jane: This work helps us get more accurate insights into how interventions affect patient survival based on individual characteristics while keeping methodological rigor high.
Lu: The paper offers a very solid foundation, paving the way for more flexible specifications in the future when we look at treatment effect heterogeneity.
Xiang Meng, Iavor Bojinov
stat.ME
Submitted: 2025-03-02
Updated: 2026-11-14
Comments: 41 pages main text, 39 pages appendix + reference, 4 figures, 13 tables
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 81/100
The gist: This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments…
Key concepts
- Staggered Adoption Designs
- This design is used when different patients receive a treatment at different times, rather than all at once. In heart transplantation, some patients get the organ earlier than others. This paper adapts this design to handle the time-varying nature of treatment assignment, which is crucial for accurate causal inference in these settings.
- Survival Analysis/Hazard Functions
- Survival analysis is a statistical method used to study the 'time until an event occurs,' such as death after a transplant. Hazard functions model the instantaneous risk of that event occurring at any given time. The paper uses these functions to incorporate the time-varying treatment timing into the causal estimation framework.
- Double Machine Learning (DML)
- DML is a technique used for robust estimation, especially when dealing with complex models or potentially incorrect assumptions about the data. It involves two stages: first, estimating propensity scores and baseline effects, and second, using those estimates to find the treatment effect. This approach helps ensure that the final causal estimate is unbiased even if some parts of the model are misspecified.
Terminology
Summary
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. p. 0 Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for.Where the paper says it
Problem Formulation and Framework
The core problem addressed is estimating causal effects in settings where treatment timing varies across patients, such as in heart transplantation where waiting times are random. Traditional causal inference fails because the treatment timing is random and highly variable, making it impossible to establish a fixed time point to separate control and treatment groups. p. 1 If there were a fixed time before which patients remained in the control group and after
which they received treatment, then we could use that fixed time to separate treatment
and control groups and conduct causal inference. However, this approach fails because
treatment timing is random and highly variable. To address this challenge, we extend staggered adoption designs to time-to-event outcomes by integrating survival analysis techniques (Cox, 1972; Klein and Moeschberger,
1997; Fleming and Harrington, 2005; Kalbfleisch and Prentice, 2011). Specifically, we use
hazard functions to model the instantaneous probability of an event. Hazard functions
allow us to characterize each unit’s contribution to the outcome at any time, regardless of
whether the unit is in its control or treatment period. We characterize a set of assumptions
under which hazard-based models can be used to solve this causal inference problem.Where the paper says it
Where the paper says it
Where the paper says it
Methodological Contributions
The paper presents two main contributions to the field. First, they bridge the gap between staggered adoption designs and survival analysis by identifying conditions under which time-varying treatment effects can be estimated in a survival framework. p. 3 The contributions of this paper are twofold. First, we bridge the gap between staggered
adoption designs and survival analysis by identifying conditions under which time-varying
treatment effects can be estimated in a survival framework. Specifically, we show how key
assumptions from staggered adoption designs can be adapted to justify the use of existing
survival analysis techniques for causal inference with time-varying treatments. Second,
we propose an estimator that addresses the complexity of real-world data, enhancing performance through Double Machine Learning (DML) techniques to ensure unbiased and
efficient estimation of treatment effects, thereby advancing survival analysis methods for
handling time-varying treatments.Where the paper says it
Where the paper says it
Estimation Procedure via DML
The authors introduce a Time-Varying Causal Survival Learner (TV-CSL) framework which is characterized by the equation νt(x) = τ (x) · at(x) + η0(x). This framework utilizes a propensity score function at time t, defined as at(x) = P(A ≤ t ∆ = 1, X), which represents the probability of adoption by time t for non-censored data. p. 9 The hazard function h(t) at time t represents the instantaneous rate of event
occurrence:
h(t) = lim
∆t→0
Pr(t ≤ T < t + ∆t | T ≥ t)
∆t
. The hazard function is particularly useful for censored data because it characterizes the
risk of an event at time t, given survival up to that time.Where the paper says it
-
First Stage (Fold One): Estimate the propensity score at(x) = P(Ai ≤ tXi = x, ∆i = 1).
-
Second Stage (Fold Two): Compute ˆνt(x) = ˆτ (x) · aˆt(x) + ˆη0(x), and then estimate the treatment effect function τ (x) = x T β by solving a minimization problem.
Performance Evaluation and Findings
Simulation studies demonstrate that the TV-CSL method consistently outperforms the S-Lasso Method across various specifications, especially when the treatment model is correctly specified. Specifically, TV-CSL maintains its advantage even when the treatment model is misspecified, attributing this robustness to the simplicity of its treatment model. p. 2 Specifically, we use
hazard functions to model the instantaneous probability of an event. Hazard functions
allow us to characterize each unit’s contribution to the outcome at any time, regardless of
whether the unit is in its control or treatment period. Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for.Where the paper says it
Where the paper says it
Limitations and Future Directions
A key limitation acknowledged is the assumption of linear treatment effect heterogeneity, suggesting future work could extend the framework to accommodate more flexible specifications of τ (x) using non-parametric or semi-parametric approaches. Another area for future research involves considering two distinct hazard functions post-transplant to capture the immediate risk increase followed by a lower risk level. The paper also suggests developing testing procedures for identifying which patients benefit most from transplantation. p. 3 The contributions of this paper are twofold. First, we bridge the gap between staggered
adoption designs and survival analysis by identifying conditions under which time-varying
treatment effects can be estimated in a survival framework.Where the paper says it
Conclusion and Implications
The proposed estimator advances the capabilities of causal inference in survival analysis by providing a robust approach for analyzing staggered treatment adoption with time-varying interventions in diverse applied contexts. The findings have implications beyond healthcare, enabling more accurate estimation of how interventions like free trials causally affect time-to-conversion outcomes in business settings. p. 3 In business settings, companies like streaming services and cloud storage providers need to understand
how their interventions affect time-dependent outcomes. For example, when companies
offer free membership trials, they must evaluate how free trial participation causally influences the time until conversion to paid memberships (Gopalakrishnan and Park, 2021;
Almathami et al. Recent work by Athey and Imbens (2022) and Goodman-Bacon (2021) has highlighted the
importance of properly accounting for treatment timing in panel data settings. Sun and
Abraham (2021) provided crucial insights into the bias that can arise from ignoring treatment effect heterogeneity across adoption cohorts. Specifically, we show how key
assumptions from staggered adoption designs can be adapted to justify the use of existing
survival analysis techniques for causal inference with time-varying treatments. Second,
we propose an estimator that addresses the complexity of real-world data, enhancing performance through Double Machine Learning (DML) techniques to ensure unbiased and
efficient estimation of treatment effects, thereby advancing survival analysis methods for
handling time-varying treatments.Where the paper says it
Where the paper says it
Where the paper says it
References
Jaap H Abbring. The event-history approach to program evaluation. p. 5 (2022) provided a comprehensive review and
tutorial on extending metalearners to handle right-censored time-to-event data. Taking a
different approach, Gao and Hastie (2021) applied the double machine learning framework
to exponential families, with Cox models serving as a specific instance. The
wide variation in treatment timing makes it impossible to establish any fixed time point
that could meaningfully separate treatment and control groups (Akintoye et al. Staggered adoption designs in econometrics (Athey and Imbens, 2022) provide a way
to estimate causal effects with random treatment times. This differs from traditional causal inference
where treatment and control groups are determined at the study’s outset and remain fixed
throughout. If there were a fixed time before which patients remained in the control group and after
which they received treatment, then we could use that fixed time to separate treatment
and control groups and conduct causal inference.Where the paper says it
Where the paper says it
Where the paper says it
Jaap H Abbring and Gerard J Van den Berg. Social experiments and instrumental variables with duration outcomes. p. 1 While this procedure is likely to extend patients’ lives on average, researchers
are particularly interested in how treatment effects vary with patient characteristics and
surgical details. Time-to-event outcomes measure how long until a critical event occurs—in
transplant cases, this is often the duration from surgery until death or organ rejection.Where the paper says it
Emmanuel Akintoye, Doosup Shin, Paulino Alvarez, and Alexandros Briasoulis. p. 2 , 1986; O’Connell et al. The
wide variation in treatment timing makes it impossible to establish any fixed time point
that could meaningfully separate treatment and control groups (Akintoye et al. After a patient is listed as a candidate for heart transplant, they must wait for an
available donor heart before undergoing the procedure (Almond et al. This means
patients who receive heart transplants experience both control time (waiting period) and
treatment time (post-transplant period).Where the paper says it
Where the paper says it
Hassan Khader Y Almathami, Khin Than Win, and Elena Vlahu-Gjorgievska. p. 3 In business settings, companies like streaming services and cloud storage providers need to understand
how their interventions affect time-dependent outcomes. For example, when companies
offer free membership trials, they must evaluate how free trial participation causally influences the time until conversion to paid memberships (Gopalakrishnan and Park, 2021;
Almathami et al. In healthcare settings, this assumption may hold when treatment decisions are based
solely on observed patient characteristics that are included in our covariate set. For heart
transplantation, this would require that all factors influencing both transplant decisions
and survival outcomes—such as disease severity, comorbidities, organ compatibility, and
functional status—are measured and accounted for in our analysis. tronic health records, but their setting differs from ours in that the treatment depends on
time-varying covariates at each point, rather than following a staggered adoption pattern. For heterogeneous treatment effect (HTE) estimation with time-to-event outcomes, researchers have developed various machine learning approaches, though none have explicitly addressed the staggered adoption design.Where the paper says it
Where the paper says it
Where the paper says it
Christopher SD Almond, Ravi R Thiagarajan, Gary E Piercey, Kimberlee Gauvreau, Elizabeth D Blume, Heather J Bastardi, Francis Fynn-Thompson, and TP Singh. Waiting list mortality among children listed for heart transplantation in the united states. p. 8 bility is often determined by allocation scores that rank patients based on observed covariates. For heart transplantation, factors such as blood type compatibility, tissue matching,
disease severity, geographical proximity to donor, and waiting time all contribute to a patient’s position on the waiting list Cascino et al.Where the paper says it
Per Kragh Andersen, Ørnulf Borgan, Richard D. Gill, and Niels Keiding. Statistical Models Based on Counting Processes. p. 2 Hazard functions
allow us to characterize each unit’s contribution to the outcome at any time, regardless of
whether the unit is in its control or treatment period. We characterize a set of assumptions
under which hazard-based models can be used to solve this causal inference problem. Imai and Kim (2021) further developed
these methods for handling staggered treatment adoption. However, despite its importance,
this crucial structure of random treatment timing remains largely unexplored in survival
analysis settings, where our work pioneers new methodological approaches.Where the paper says it
Where the paper says it
Susan Athey and Guido W Imbens. p. 2 The
wide variation in treatment timing makes it impossible to establish any fixed time point
that could meaningfully separate treatment and control groups (Akintoye et al. Staggered adoption designs in econometrics (Athey and Imbens, 2022) provide a way
to estimate causal effects with random treatment times. The challenges of treatments with random timing have been extensively studied in the
econometrics literature, particularly in the context of difference-in-differences designs. Recent work by Athey and Imbens (2022) and Goodman-Bacon (2021) has highlighted the
importance of properly accounting for treatment timing in panel data settings. The
wide variation in treatment timing makes it impossible to establish any fixed time point
that could meaningfully separate treatment and control groups (Akintoye et al. Staggered adoption designs in econometrics (Athey and Imbens, 2022) provide a way
to estimate causal effects with random treatment times.Where the paper says it
Where the paper says it
Where the paper says it
Thomas M. Cascino, Jeffrey S. p. 8 bility is often determined by allocation scores that rank patients based on observed covariates. For heart transplantation, factors such as blood type compatibility, tissue matching,
disease severity, geographical proximity to donor, and waiting time all contribute to a patient’s position on the waiting list Cascino et al. (1992) and Hern’an (2010) introduced key methodological frameworks
for causal inference in survival settings. Li and Greene (2015) developed doubly robust
estimators for survival outcomes, while Zhang et al. Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. , 2021; DeFilippis et al. , 2022)
However, measuring the causal effect of such medical interventions is not straightforward.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. Double/debiased machine learning for treatment and structural parameters, 2018. p. 2 To handle these complexities, we employ double machine learning
(DML) techniques (Chernozhukov et al. , 2018; K¨unzel et al.Where the paper says it
Swati Choudhry, Yunfei Wang, Susan W Denfield, Antonio G Cabrera, Jack F Price, Hari P Tunuguntla, Vikas R Dharnidharka, Charles E Canter, and William J Dreyer. A recipient risk prediction tool for short-term mortality after pediatric heart transplantation. p. 0 Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. , 2018; K¨unzel et al. , 2019; Nie and Wager, 2021;
Gao and Hastie, 2021) in our estimation procedure.Where the paper says it
Where the paper says it
David R Cox. p. 2 , to study a policy’s effect on a
country’s GDP, you obtain the country’s GDP over many years, with some observations
before the policy and some after), they cannot handle time-to-event outcomes that are
observed only once—either at the event occurrence (such as mortality) or at study end. To address this challenge, we extend staggered adoption designs to time-to-event outcomes by integrating survival analysis techniques (Cox, 1972; Klein and Moeschberger,
1997; Fleming and Harrington, 2005; Kalbfleisch and Prentice, 2011). Hazard functions
allow us to characterize each unit’s contribution to the outcome at any time, regardless of
whether the unit is in its control or treatment period. We characterize a set of assumptions
under which hazard-based models can be used to solve this causal inference problem. Despite these developments, the literature lacks
methods specifically designed for causal inference in time-to-event settings with variable
treatment timing as encountered in staggered adoption designs. The organization of the paper is as follows: In Section 2, we formalize the notation,
introduce assumptions for the causal framework, and present the statistical problem.Where the paper says it
Where the paper says it
Where the paper says it
John Crowley and Marie Hu. p. 1 1 Introduction
In healthcare, understanding the causal effect of medical interventions on patient survival
is crucial. Heart transplantation is a compelling example, as demonstrated by the Stanford
Heart Transplant study (Crowley and Hu, 1977; Zhu et al. Furthermore, we enhance estimation performance by incorporating double machine learning
methods, improving efficiency when handling complex relationships between patient
characteristics and survival outcomes. Through both simulation studies and application to heart transplant data, our approach demonstrates superior performance
compared to traditional methods, reducing bias and offering theoretical guarantees
for improved efficiency in survival analysis settings. 1 Introduction
In healthcare, understanding the causal effect of medical interventions on patient survival
is crucial. Heart transplantation is a compelling example, as demonstrated by the Stanford
Heart Transplant study (Crowley and Hu, 1977; Zhu et al.Where the paper says it
Where the paper says it
Where the paper says it
Yifan Cui, Michael R Kosorok, Erik Sverdrup, Stefan Wager, and Ruoqing Zhu. p. 2 , 2018; K¨unzel et al. , 2019; Nie and Wager, 2021;
Gao and Hastie, 2021) in our estimation procedure. For heterogeneous treatment effect (HTE) estimation with time-to-event outcomes, researchers have developed various machine learning approaches, though none have explicitly addressed the staggered adoption design. (2023) developed causal survival
forests that adjust for right censored time-to-event outcome using doubly robust estimating equations, while Xu et al. Despite these developments, the literature lacks
methods specifically designed for causal inference in time-to-event settings with variable
treatment timing as encountered in staggered adoption designs. The organization of the paper is as follows: In Section 2, we formalize the notation,
introduce assumptions for the causal framework, and present the statistical problem.Where the paper says it
Where the paper says it
Where the paper says it
Ersilia M DeFilippis, Kiran K Khush, Maryjane A Farr, Amy Fiedler, Arman Kilic, and Michael M Givertz. p. 1 , 2007; Kilic et al. , 2021; DeFilippis et al. In healthcare settings, this assumption may hold when treatment decisions are based
solely on observed patient characteristics that are included in our covariate set. For heart
transplantation, this would require that all factors influencing both transplant decisions
and survival outcomes—such as disease severity, comorbidities, organ compatibility, and
functional status—are measured and accounted for in our analysis. , 2021; DeFilippis et al. , 2022)
However, measuring the causal effect of such medical interventions is not straightforward.Where the paper says it
Where the paper says it
Where the paper says it
Oliver Dukes, Mats J Stensrud, Riccardo Brioschi, and Aaron Hudson. Nonparametric tests of treatment effect homogeneity for policy-makers. p. 2 While these designs work well for repeatedly measured continuous outcomes,
where a unit’s outcome is obtained at different times (e. , to study a policy’s effect on a
country’s GDP, you obtain the country’s GDP over many years, with some observations
before the policy and some after), they cannot handle time-to-event outcomes that are
observed only once—either at the event occurrence (such as mortality) or at study end. Through both simulation studies and application to heart transplant data, our approach demonstrates superior performance
compared to traditional methods, reducing bias and offering theoretical guarantees
for improved efficiency in survival analysis settings. 1
arXiv:2503. (2023) developed causal survival
forests that adjust for right censored time-to-event outcome using doubly robust estimating equations, while Xu et al. (2024) proposed censoring unbiased transformations with
orthogonality properties that can be integrated with existing HTE learners in scenarios
with or without competing risks.Where the paper says it
Where the paper says it
Where the paper says it
Roger W Evans, Diane L Manninen, Louis P Garrison, and Anthony M Maier. p. 2 ranges from a few days to more than a year (Evans et al. , 1986; O’Connell et al. Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. ranges from a few days to more than a year (Evans et al. , 1986; O’Connell et al.Where the paper says it
Where the paper says it
Where the paper says it
Alyssa Power, Kurt R. p. 8 (2013); Parker et al. (2019); Power et al. Consider i = 1, . Let Ti ∈ [0,∞) denote the time until an event of interest
occurs, such as time until mortality in transplant studies. (2023) developed causal survival
forests that adjust for right censored time-to-event outcome using doubly robust estimating equations, while Xu et al. (2024) proposed censoring unbiased transformations with
orthogonality properties that can be integrated with existing HTE learners in scenarios
with or without competing risks.Where the paper says it
Where the paper says it
Where the paper says it
James M Robins, Andrea Rotnitzky, and Lue Ping Zhao. p. 4 However, these approaches primarily focus on static treatments or simple time-varying confounders, rather than addressing the complexities of variable treatment timing. For example, Robins et al. Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. For example, Robins et al. (1992) and Hern’an (2010) introduced key methodological frameworks
for causal inference in survival settings.Where the paper says it
Where the paper says it
Where the paper says it
Donald B Rubin. p. 6 We also observe pre-treatment covariates Xi ∈ R
p
. We adapt the following standard
causal assumptions from Rubin (1974):
Assumption 1 (Stable Unit Treatment Value Assumption, SUTVA): Each unit’s
potential outcome is determined solely by its own treatment assignment, with no interference between units and uniform treatment versions. For example, when companies
offer free membership trials, they must evaluate how free trial participation causally influences the time until conversion to paid memberships (Gopalakrishnan and Park, 2021;
Almathami et al. As in transplant studies, both the timing of the intervention
(free trial start) and the outcome (conversion) vary across units, making traditional causal
inference methods insufficient for estimating these effects. We also observe pre-treatment covariates Xi ∈ R
p
. We adapt the following standard
causal assumptions from Rubin (1974):
Assumption 1 (Stable Unit Treatment Value Assumption, SUTVA): Each unit’s
potential outcome is determined solely by its own treatment assignment, with no interference between units and uniform treatment versions.Where the paper says it
Where the paper says it
Where the paper says it
Azeem M Shaikh and Panos Toulis. p. 4 Our work employs Cox models for treatment modeling and focuses on learning the hazard ratio for survival outcomes, and this is different from other recent work on Cox models
in causal inference. In related work, Shaikh and Toulis (2021) developed a framework
for testing treatment effects in time-varying settings using Cox models, but they concentrate on testing the null hypothesis of no treatment effect on continuous outcomes. Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. Despite these developments, the literature lacks
methods specifically designed for causal inference in time-to-event settings with variable
treatment timing as encountered in staggered adoption designs. The organization of the paper is as follows: In Section 2, we formalize the notation,
introduce assumptions for the causal framework, and present the statistical problem.Where the paper says it
Where the paper says it
Where the paper says it
Liyang Sun and Sarah Abraham. p. 4 Recent work by Athey and Imbens (2022) and Goodman-Bacon (2021) has highlighted the
importance of properly accounting for treatment timing in panel data settings. Sun and
Abraham (2021) provided crucial insights into the bias that can arise from ignoring treatment effect heterogeneity across adoption cohorts. Time-to-event outcomes measure how long until a critical event occurs—in
transplant cases, this is often the duration from surgery until death or organ rejection. Understanding these heterogeneous treatment effects—how the impact varies as a function
of patient characteristics—is essential for improving patient selection criteria, optimizing
intervention timing, and ultimately enhancing survival outcomes in transplant medicine
(Trulock et al. The challenges of treatments with random timing have been extensively studied in the
econometrics literature, particularly in the context of difference-in-differences designs. Recent work by Athey and Imbens (2022) and Goodman-Bacon (2021) has highlighted the
importance of properly accounting for treatment timing in panel data settings.Where the paper says it
Where the paper says it
Where the paper says it
J. Kenneth Tay, Balasubramanian Narasimhan, and Trevor Hastie. p. 2 We characterize a set of assumptions
under which hazard-based models can be used to solve this causal inference problem. In addition to handling random treatment timing, we address the complex, non-linear
relationships between covariates and outcomes that often arise in real-world applications
(Hastie et al. We characterize a set of assumptions
under which hazard-based models can be used to solve this causal inference problem. In addition to handling random treatment timing, we address the complex, non-linear
relationships between covariates and outcomes that often arise in real-world applications
(Hastie et al. For heterogeneous treatment effect (HTE) estimation with time-to-event outcomes, researchers have developed various machine learning approaches, though none have explicitly addressed the staggered adoption design. (2023) developed causal survival
forests that adjust for right censored time-to-event outcome using doubly robust estimating equations, while Xu et al.Where the paper says it
Where the paper says it
Where the paper says it
Elbert P Trulock, Jason D Christie, Leah B Edwards, Mark M Boucek, Paul Aurora, David O Taylor: Towards optimal doubly robust estimation of heterogeneous causal effects, August 2023. p. 5 For heterogeneous treatment effect (HTE) estimation with time-to-event outcomes, researchers have developed various machine learning approaches, though none have explicitly addressed the staggered adoption design. (2023) developed causal survival
forests that adjust for right censored time-to-event outcome using doubly robust estimating equations, while Xu et al.Where the paper says it
Hertz. Registry of the international society for heart and lung transplantation: twentyfourth official adult lung and heart–lung transplantation report—2007. p. 0 Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for.Where the paper says it
Where the paper says it
Aad W Van der Vaart. Asymptotic statistics, volume 3. Cambridge university press, 2000.
Stijn Vansteelandt and Marshall Joffe. p. 4 (2017) proposed methods for handling time-dependent confounding. Vansteelandt and Joffe (2014) considered time-varying
treatments, but in their work, treatment timing is well-defined, such as scheduled half-year
visits for people with HIV, and treatment can be reversed. Hazard functions
allow us to characterize each unit’s contribution to the outcome at any time, regardless of
whether the unit is in its control or treatment period. We characterize a set of assumptions
under which hazard-based models can be used to solve this causal inference problem. (2017) proposed methods for handling time-dependent confounding. Vansteelandt and Joffe (2014) considered time-varying
treatments, but in their work, treatment timing is well-defined, such as scheduled half-year
visits for people with HIV, and treatment can be reversed.Where the paper says it
Where the paper says it
Where the paper says it
Yuyao Wang, Andrew Ying, and Ronghui Xu. p. 5 Taking a
different approach, Gao and Hastie (2021) applied the double machine learning framework
to exponential families, with Cox models serving as a specific instance. More recently,
Wang et al. (2024) addressed estimation for left-truncated right-censored (LTRC) data,
a significant advancement in the field. Despite these developments, the literature lacks
methods specifically designed for causal inference in time-to-event settings with variable
treatment timing as encountered in staggered adoption designs.Where the paper says it
Where the paper says it
Shenbo Xu, Raluca Cobzaru, Stan N. Finkelstein, Roy E. Welsch, Kenney Ng, and Zach Shahn. Estimating Heterogeneous Treatment Effects on Survival Outcomes Using Counterfactual Censoring Unbiased Transformations, September 2024. p. 5 (2023) developed causal survival
forests that adjust for right censored time-to-event outcome using doubly robust estimating equations, while Xu et al. (2024) proposed censoring unbiased transformations with
orthogonality properties that can be integrated with existing HTE learners in scenarios
with or without competing risks.Where the paper says it
Yizhe Xu, Nikolaos Ignatiadis, Erik Sverdrup, Scott Fleming, Stefan Wager, and Nigam Shah. Treatment Heterogeneity for Survival Outcomes, September 2022. p. 4 Recent work by Athey and Imbens (2022) and Goodman-Bacon (2021) has highlighted the
importance of properly accounting for treatment timing in panel data settings. Sun and
Abraham (2021) provided crucial insights into the bias that can arise from ignoring treatment effect heterogeneity across adoption cohorts.Where the paper says it
Ping Zhang, Fei Wang, Jianying Hu, and Robert Sorrentino. p. 4 (1992) and Hern’an (2010) introduced key methodological frameworks
for causal inference in survival settings. Li and Greene (2015) developed doubly robust
estimators for survival outcomes, while Zhang et al. everyone receives the same version of treatment in healthcare settings, which might not hold
in practice—for example, surgical procedures may vary by surgeon expertise or hospital
resources—so caution is warranted when applying these methods in real-world clinical
scenarios. Assumption 2 (Unconfoundedness): The treatment assignment is unconfounded, conditional on covariates Xi
. Li and Greene (2015) developed doubly robust
estimators for survival outcomes, while Zhang et al. (2017) proposed methods for handling time-dependent confounding.Where the paper says it
Where the paper says it
Where the paper says it
Jie Zhu and Blanca Gallego. p. 4 In related work, Shaikh and Toulis (2021) developed a framework
for testing treatment effects in time-varying settings using Cox models, but they concentrate on testing the null hypothesis of no treatment effect on continuous outcomes. In the
biomedical domain, Zhu and Gallego (2022) considered time-varying treatments in elec5 Our work employs Cox models for treatment modeling and focuses on learning the hazard ratio for survival outcomes, and this is different from other recent work on Cox models
in causal inference. In related work, Shaikh and Toulis (2021) developed a framework
for testing treatment effects in time-varying settings using Cox models, but they concentrate on testing the null hypothesis of no treatment effect on continuous outcomes. In related work, Shaikh and Toulis (2021) developed a framework
for testing treatment effects in time-varying settings using Cox models, but they concentrate on testing the null hypothesis of no treatment effect on continuous outcomes. In the
biomedical domain, Zhu and Gallego (2022) considered time-varying treatments in elec5Where the paper says it
Where the paper says it
Where the paper says it
Yuanjia Zhu, Bharathi Lingala, Michael Baiocchi, Veronica Toro Arana, Kiah M Williams, Yasuhiro Shudo, Philip E Oyer, and Y Joseph Woo. The stanford experience of heart transplantation over five decades. p. 0 Time-Varying Causal Survival Learning
Xiang Meng, Iavor Bojinov
March 4, 2025
Abstract
This work bridges the gap between staggered adoption designs and survival analysis to estimate causal effects in settings with time-varying treatments, addressing a
fundamental challenge in medical research exemplified by the Stanford Heart Transplant study. In medical interventions, particularly organ transplantation, the timing
of treatment varies significantly across patients due to factors such as donor availability and patient readiness, introducing potential bias in treatment effect estimation if
not properly accounted for. Heart transplantation is a compelling example, as demonstrated by the Stanford
Heart Transplant study (Crowley and Hu, 1977; Zhu et al. , 2021), where patients with
end-stage heart failure undergo surgery to replace their failing hearts with healthy donor
hearts.Where the paper says it
Where the paper says it
A Derivation of the TV-CSL Estimator. p. 3 Specifically, we show how key
assumptions from staggered adoption designs can be adapted to justify the use of existing
survival analysis techniques for causal inference with time-varying treatments. Second,
we propose an estimator that addresses the complexity of real-world data, enhancing performance through Double Machine Learning (DML) techniques to ensure unbiased and
efficient estimation of treatment effects, thereby advancing survival analysis methods for
handling time-varying treatments.Where the paper says it
B Key Lemma for Deriving the Score Function. p. 3 First, we bridge the gap between staggered
adoption designs and survival analysis by identifying conditions under which time-varying
treatment effects can be estimated in a survival framework. Specifically, we show how key
assumptions from staggered adoption designs can be adapted to justify the use of existing
survival analysis techniques for causal inference with time-varying treatments. Specifically, we show how key
assumptions from staggered adoption designs can be adapted to justify the use of existing
survival analysis techniques for causal inference with time-varying treatments. Second,
we propose an estimator that addresses the complexity of real-world data, enhancing performance through Double Machine Learning (DML) techniques to ensure unbiased and
efficient estimation of treatment effects, thereby advancing survival analysis methods for
handling time-varying treatments.Where the paper says it
Where the paper says it
Plugin r:= (Y − ψ′(η)) /ψ′′(η), and taking x, w as random variables, we have E [(W − a(X)) (Y − (ψ′(W η1(X) + (1 − W)η0(X)))! p. 6 2 Problem Set Up
2. 1 Notation and data
Let capital letters denote random variables and lowercase letters denote their realizations.Where the paper says it
Proof of Proposition 1. We write γn(x) = γ(x) + αnξn(x), where E [ξ 2 n (X)] = 1 is a unit directional vector, and αn is the distance from γ n(x) to γ(x). Similarly, we can write η n(x) = η(x) + ρ nζ n(x), where E [ζ 2 n (X)] = 1.
β n − β0 = Op n-1/2 + α 2 n + ρ 2 n.
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
-
Use Time-Varying Causal Survival Learner to estimate causal effects in settings with time-varying treatments by bridging staggered adoption designs and survival analysis techniques. This allows for causal inference when treatment timing varies randomly, such as in heart transplant studies.
-
Employ a double machine learning framework, specifically TV-CSL, to provide robust estimation of heterogeneous treatment effects even when the baseline log hazard function is complex or potentially misspecified. This addresses the vulnerability of outcome-based methods to baseline hazard misspecification.
-
Implement a time-varying causal survival learner that models the treatment effect as at(x) τ(x) + η0(x), where at(x) is the probability of adoption by time t and η0(x) is the baseline log hazard function. This captures the dynamic nature of treatment assignment over time.
-
Utilize a two-stage cross-fitting procedure in the TV-CSL method, where the first stage estimates at(x) under both correct and misspecified settings, followed by an outcome model estimation using Lasso with specifications matching S-Lasso. This ensures double robustness and improved convergence rates.
-
Apply the results to clinical decision support systems for organ transplantation to better understand how interventions like heart transplants affect patient survival based on individual characteristics, helping improve patient selection and timing of procedures.
Abstract
Staggered adoptions are settings where study objects convert from control to treatment at different times. Often the outcome is a time to an event, such as survival after an organ transplant or the time until an employee becomes a regular user of a new AI tool. Although many papers have studied staggered adoptions on repeatedly measured outcomes, little work has addressed survival outcomes, which are observed at most once and may be censored. We give causal assumptions under which a time-dependent survival model identifies a heterogeneous treatment effect, which tells who benefits most. Because a direct Cox learner is biased when baseline risk is learned flexibly, we introduce a time-varying, event-risk propensity score. Constructing this propensity score is nontrivial because both treatment and outcome are times to events, so it must track which individuals have not yet experienced the event and the treatment status of each at the time. Unlike existing propensity scores, centering by this propensity score makes the Cox partial-likelihood equation Neyman orthogonal. We establish the double machine learning property and the corresponding convergence rate, and an influence-function-based variance estimator. Simulations show that orthogonalization reduces error and supports near-nominal inference when the baseline risk is complex and the samples are large, whereas direct Cox estimation is more precise when baseline risk is linear. We illustrate the framework using lawyer-level Copilot adoption data and the Stanford Heart Transplant study.
Sources
- Nonparametric tests of treatment effect homogeneity for policy-makers
- Orthogonal Statistical Learning
- Estimating Heterogeneous Treatment Effects for General Responses
- Towards optimal doubly robust estimation of heterogeneous causal effects
- A Liberating Framework from Truncation and Censoring, with Application to Learning Treatment Effects
- Estimating Heterogeneous Treatment Effects on Survival Outcomes Using Counterfactual Censoring Unbiased Transformations
- Treatment Heterogeneity for Survival Outcomes
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes