Estimating Treatment Effects Through Intermediate-Event Pathways with Multi-State Data
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: "Estimating Treatment Effects Through Intermediate-Event Pathways with Multi-State Data".
Jane: Here's the gist of the paper. Estimating Pathway Treatment Effects in the Presence of Intermediate Events with Multi-State Data is a framework designed to estimate pathway treatment effects by considering hypothetical interventions in transitions between event statuses to mimic the treatment mechanism, which enables the examination of treatment effects through each transition, on each event, and along each path The Core Problem and Proposed Solution When evaluating the effect of a new drug on a survival endpoint, individuals may experience multiple intermediate events, thereby generating multiple transition pathways from baseline to the primary endpoint To understand the effect along each pathway, we consider hypothetical interventions in transitions between event statuses to mimic the treatment mechanism This framework constructs multiply robust and semiparametrically efficient estimators for pathway treatment effects based on derived efficient influence functions for counterfactual cumulative incidences under hypothetical interventions.
Tom: Framework Development The authors extend the randomized (stochastic) interventional effects framework, originally proposed for mediation analysis in longitudinal studies, to accommodate multi-state data The randomized intervention envisions random draws of the intermediate event process from a specific distribution, either conditioning on or marginalizing over time-varying confounding.
Jane: The primary object of interest is the counterfactual cumulative incidence function (CIF) of state j in G, given by F a j(t):= P(T a j ≤ t) = P(j ∈ X a(t)), t ∈
0, t∗: .
Tom: The counterfactual CIF can be decomposed as F a j(t) = X q∈Qj F a j(t; q), where Qkg represents the set of paths from state k to state g.
Paper summary: Tom: So to wrap up, the title "Estimating Treatment Effects Through Intermediate-Event Pathways with Multi-State Data" is about moving beyond simple cause and effect when multiple things happen in between. The authors claim they can estimate treatment effects by considering hypothetical interventions at each transition to see how it affects every possible pathway from start to finish.
Jane: It’s important because it shows that when we have multi-state data, which is common in survival studies, we can get more nuanced answers about a drug's effect on specific routes of disease progression.
Lu: The implication is that instead of just getting one big number for the total effect, researchers can pinpoint exactly which part of the treatment mechanism is driving which outcome along a particular sequence of events.
Meng: This gives us better targets for precision medicine because we can identify pathways where the treatment is helping, and pathways where it's failing on a specific step.
Jane: It means that for complex treatments, understanding the path-specific effects is crucial for making informed decisions about which post-treatment interventions to recommend.
Conclusion: Tom: So, this paper is about estimating how a treatment works when there are multiple steps or events happening in between.
Jane: It’s titled "Estimating Treatment Effects Through Intermediate-Event Pathways with Multi-State Data," and it looks like the authors are using hypothetical interventions to see how the drug affects each transition.
Lu: They're building on an existing framework for mediation analysis, but they're adapting it for data with multiple event statuses.
Meng: So, instead of just looking at the total effect of a drug on a final outcome, this method lets you check if the treatment works differently at each stage of the disease process.
Tom: Exactly. It allows you to see if the treatment helps one specific step but not another along a given pathway to that final result.
Jane: The authors found they can get these path-specific effects, and they used this efficient influence function method, which means their estimates are quite precise even with all the uncertainty in multi-state data.
Lu: The core idea is that you decompose the overall effect into effects from intervening at each transition, which lets you isolate those specific pathway contributions.
Meng: For us in production, that’s useful because if we know a treatment works best at step two but not step three, we can tailor the drug delivery or patient monitoring accordingly.
Tom: It moves beyond just saying "the drug works" to explaining "how and where it works."
Jane: And the implications are big for personalized medicine because understanding these pathway effects helps us predict which patients will respond best to a specific intervention at different points in their recovery.
Lu: The real power here is that they provide a method that doesn't require you to model all the complex confounding factors caused by the treatment itself, which makes it more robust.
Meng: But what about when the data has a lot of missing information or lots of time-varying things mixed in?
Tom: That’s actually one of their future directions, where they might look at intervening in event counting processes conditional on those time-varying variables.
Jane: Right, so while this framework gives us a solid foundation for understanding the "how" and "where" of treatment effects now, the next big challenge is making it work cleanly when the patient’s health situation changes constantly over time.
Fred Hutchinson Cancer Center · University of California San Diego · University of Michigan
stat.ME, econ.EM, math.ST, stat.TH
Submitted: 2026-08-23
Updated: 2026-11-05
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Estimating Pathway Treatment Effects in the Presence of Intermediate Events with Multi-State Data is a framework designed to estimate pathway treatment effects by considering hypothetical…
Key concepts
- Pathway Treatment Effects
- This refers to measuring how a treatment affects an outcome by looking at the specific sequence of events that occur over time. Instead of just looking at the final outcome, this method examines the effect along different routes or paths an individual takes from their starting point to the end.
- Intermediate Events
- These are events that happen between a person's baseline and their primary endpoint. In survival analysis, these can be various medical events (like a cardiovascular event) that occur before the final outcome is reached, creating multiple potential paths.
- Counterfactual Cumulative Incidence Function (CIF)
- This function estimates the probability of reaching a specific state or event by a certain time under different conditions. The paper uses it to model what would happen to an individual's health status if they had received the treatment versus not receiving it, considering all possible intermediate steps.
Terminology
Summary
Estimating Pathway Treatment Effects in the Presence of Intermediate Events with Multi-State Data is a framework designed to estimate pathway treatment effects by considering hypothetical interventions in transitions between event statuses to mimic the treatment mechanism, which enables the examination of treatment effects through each transition, on each event, and along each path
The Core Problem and Proposed Solution
When evaluating the effect of a new drug on a survival endpoint, individuals may experience multiple intermediate events, thereby generating multiple transition pathways from baseline to the primary endpoint To understand the effect along each pathway, we consider hypothetical interventions in transitions between event statuses to mimic the treatment mechanism This framework constructs multiply robust and semiparametrically efficient estimators for pathway treatment effects based on derived efficient influence functions for counterfactual cumulative incidences under hypothetical interventions.
Framework Development
The authors extend the randomized (stochastic) interventional effects framework, originally proposed for mediation analysis in longitudinal studies, to accommodate multi-state data The randomized intervention envisions random draws of the intermediate event process from a specific distribution, either conditioning on or marginalizing over time-varying confounding. The primary object of interest is the counterfactual cumulative incidence function (CIF) of state j in G, given by F a j(t):= P(T a j ≤ t) = P(j ∈ X a(t)), t ∈ [0, t∗]. p. 7 (2)
The primary object of interest is the time to reach a specific state j ∈ G (typically a terminal event
such as death), defined as T
a
j = inf{t ∈ [0, t∗
] : Xa
(t) = j} if state j is on the trajectory X
a
(t
∗
),
otherwise we set T
a
j = ∞. The primary estimand is the counterfactual cumulative incidence
function (CIF) of state j ∈ G, given by
F
a
j
(t) := P(T
a
j ≤ t) = P(j ∈ X
a
(t)), t ∈ [0, t∗
]. Let Gk
represent the set of states with one-step transitions from state k to its adjacent
following states. We denote the set of paths from state k to state g as Qkg, and the set of paths
5Where the paper says it
Where the paper says it
Identification and Estimation
The identification of the counterfactual cumulative incidence F a j(t) is presented in Theorem 1 under Assumptions 1–4, which ensure that potential transition hazards can be identified as observed hazards. The total effect on an event can be decomposed into the sum of interventional effects arising from sequentially intervening at each transition. p. 4 We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. The total effect on an event can be decomposed into the sum of interventional effects arising
from sequentially intervening at each transition. We derive the efficient influence function for
the counterfactual cumulative incidence under semiparametric theory (Bickel et al. 1993) and
provide a semiparametrically efficient, multiply robust estimator that does not require modeling of
treatment-induced confounding.Where the paper says it
Where the paper says it
Application and Findings
Analyzing data from the LEADER Trial, liraglutide significantly reduces the risk of cardiovascular and microvascular events The reduction in all-cause mortality is primarily mediated by its effects on expanded major adverse cardiovascular events. Path-specific effects on death are significant, such as the effect along the path O→E→D with a restricted mean survival time gained of −0. p. 3 Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect. Estimating transition-, event-, and path-specific effects generally concerns estimating
counterfactual cumulative incidences under hypothetical interventions imposed on transitions.Where the paper says it
Robustness and Conclusions
The EIF-based estimator is shown to be consistent if, for each q ∈ Qj, at least one of several conditions holds regarding the specification of transition hazards, propensity score, and censoring hazard. The EIF-based estimator exhibits multiple robustness and semiparametric efficiency as established in Theorem 3. p. 4 We derive the efficient influence function for
the counterfactual cumulative incidence under semiparametric theory (Bickel et al. 1993) and
provide a semiparametrically efficient, multiply robust estimator that does not require modeling of
treatment-induced confounding. By analyzing data from the LEADER Trial, we
find that liraglutide significantly reduces the risk of cardiovascular and microvascular events. The reduction in all-cause mortality is primarily mediated by its effects on expanded major
adverse cardiovascular events.Where the paper says it
Where the paper says it
Future Directions
Alternative interventions may be considered when time-varying confounders are present, such as intervening in event-counting processes conditional on time-varying covariates. The separable effects framework assumes that the initial treatment can be decomposed into components, each exerting a direct effect on a specific event or a pathway, although identifying clinically meaningful components with isolated effects is not always feasible. p. 2 Separable effects target the effect of each
component, but are not informative about the effect through each transition pathway. Even if
the effects of each component are theoretically well-defined, it is hard to identify biologically
meaningful treatment components with isolated effects in a real trial. The randomized intervention envisions
random draws of the intermediate event process from a specific distribution, either conditioning
on or marginalizing over time-varying confounding (Deng et al. When unmeasured
confounders are present, the randomized intervention remains valid in a hypothetical sense by
marginalizing over them.Where the paper says it
Where the paper says it
Data Availability
The data that support the findings of this study are available from Novo Nordisk A/S
Supplementary Material Details
The online Supplementary Material includes regularity conditions, details of sensitivity analysis, simulation studies, additional data analysis results, and proofs of theoretical results. The EIF-based estimator is given by the formula in (S1). p. 8 graph of the counterfactual event-counting processes: Na
E
(t) for E, Na
M(t) for M, and Na
D
(t) for D. Given the history up to time t, if the individual is in the initial state O, we draw Na
OE(t) based on
the transition hazard of O→E under treated, Na
OM(t) based on the transition hazard of O→M under
control, and Na
OD(t) based on the transition hazard of O→D under control; if the individual is in
state E, we draw Na
EM(t) based on the transition hazard of E→M under control and Na
ED(t) based on
the transition hazard of E→D under treated; if the individual is in state M, we draw Na
ME(t) based
on the transition hazard of M→E under control and Na
MD(t) based on the transition hazard of M→D
under control. We rigorously
specify the sequential randomized interventions in event-counting processes that control transition
hazards by marginalizing over potential treatment-induced confounding. We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. Identifying treatment effects under hypothetical interventions does not require
cross-world assumptions such as sequential ignorability as long as the hypothetical intervention is
meaningful. The randomized interventional effects framework is powerful in studies with multiple
3 By analyzing data from the LEADER Trial, we
find that liraglutide significantly reduces the risk of cardiovascular and microvascular events. The reduction in all-cause mortality is primarily mediated by its effects on expanded major
adverse cardiovascular events. We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. The total effect on an event can be decomposed into the sum of interventional effects arising
from sequentially intervening at each transition. We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. The total effect on an event can be decomposed into the sum of interventional effects arising
from sequentially intervening at each transition. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. By analyzing data from the LEADER Trial, we
find that liraglutide significantly reduces the risk of cardiovascular and microvascular events. The reduction in all-cause mortality is primarily mediated by its effects on expanded major
adverse cardiovascular events. The effects operating through each pathway provide insight into multiple
questions, such as how the drug is effective and whether additional post-treatment interventions
should be recommended. For example, to study the direct effect, researchers wish to isolate
the effect of treatment itself on the primary endpoint, independent of any modification through
intermediate events. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. The effects operating through each pathway provide insight into multiple
questions, such as how the drug is effective and whether additional post-treatment interventions
should be recommended. For example, to study the direct effect, researchers wish to isolate
the effect of treatment itself on the primary endpoint, independent of any modification through
intermediate events. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. The effects operating through each pathway provide insight into multiple
questions, such as how the drug is effective and whether additional post-treatment interventions
should be recommended. For example, to study the direct effect, researchers wish to isolate
the effect of treatment itself on the primary endpoint, independent of any modification through
intermediate events. The effects operating through each pathway provide insight into multiple
questions, such as how the drug is effective and whether additional post-treatment interventions
should be recommended. For example, to study the direct effect, researchers wish to isolate
the effect of treatment itself on the primary endpoint, independent of any modification through
intermediate events. The effects operating through each pathway provide insight into multiple
questions, such as how the drug is effective and whether additional post-treatment interventions
should be recommended. For example, to study the direct effect, researchers wish to isolate
the effect of treatment itself on the primary endpoint, independent of any modification through
intermediate events. We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. The total effect on an event can be decomposed into the sum of interventional effects arising
from sequentially intervening at each transition. Estimating Pathway Treatment Effects in the
Presence of Intermediate Events with Multi-State
Data
Yuhao Deng*
Vaccine and Infectious Disease Division, Fred Hutchinson Cancer Center
Haoyu Wei∗
Department of Economics, University of California San Diego
Donglin Zeng
Department of Biostatistics, University of Michigan
Rui Song
Amazon Inc. Xiao-Hua Zhou
Department of Biostatistics and Beijing International Center for
Mathematical Research, Peking University
Abstract
During clinical trials evaluating a drug’s effect on a survival endpoint, intermediate events
often occur in addition to the primary event. We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. The total effect on an event can be decomposed into the sum of interventional effects arising
from sequentially intervening at each transition. We demonstrate the
identifiability of the counterfactual cumulative incidence of each event under any given intervention. The total effect on an event can be decomposed into the sum of interventional effects arising
from sequentially intervening at each transition.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Table S2 Summary
Table S2 lists the estimated coefficients with standard errors in four Cox models for non-fatal EMACE, MVE, direct CVD, and other-cause death. The first row of Figure S2 shows the estimated cumulative incidence functions of EMACE, MVE, and death. p. 5 We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment.Where the paper says it
Where the paper says it
Table S1 Summary
Table S1 summarizes the average bias, standard deviation (SD), average standard error (SE), and coverage rate (CR) of 95% nominal confidence intervals of the EIF-based estimator (Eff) and plug-in estimator (Reg) for F a j(t) at selected time points. The CR for the EIF-based estimator is close to 95%. p. 0 The
hypothetical interventions adjust for effects through intermediate events and marginalize over
unobserved treatment-induced confounding, if any. Based on the derived efficient influence
functions for the counterfactual cumulative incidences under hypothetical interventions, we
construct multiply robust and semiparametrically efficient estimators for pathway treatment
effects.Where the paper says it
Figure S3 Summary
Figure S3 shows the estimated path-specific effects on death by intervening in each path from the initial state to death. The effect along the path O→D is −0. p. 3 Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment. Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect. Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment. Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect.Where the paper says it
Where the paper says it
Figure S4 Summary
Figure S4 presents the counterfactual CIFs of MACE, MVE, and death under all possible dynamic treatment policies. The difference between O1E0 and O0E1 yields a P-value of 0. p. 1 By identifying the transition pathways through which treatment is ineffective or
harmful, doctors can impose additional medications to prevent these transitions. Treating the primary event and intermediate events as distinct states yields the so-called multi-state
data (Hougaard 1999, Andersen & Keiding 2002, Putter et al. O D
E
M
O D
E
M
O D
E
M
O D
E
M
O D
E
M
q1 = O→D q2 = O→E→D
q3 = O→M→D q4 = O→E→M→D q5 = O→M→E→D
Figure 1: Five potential paths from baseline (O) to death (D) with expanded major adverse
cardiovascular events (E) and microvascular events (M) as two intermediate states. from initial states to state g as Qg. By identifying the transition pathways through which treatment is ineffective or
harmful, doctors can impose additional medications to prevent these transitions. Treating the primary event and intermediate events as distinct states yields the so-called multi-state
data (Hougaard 1999, Andersen & Keiding 2002, Putter et al.Where the paper says it
Where the paper says it
Where the paper says it
Figure S5 Summary
Figure S5 shows the estimated direct treatment effect on EMACE, MVE, and other-cause death by intervening in the cause-specific hazards of the event of interest (with frailty). Liraglutide reduces the risks of EMACE and MVE. p. 0 Our proposed framework enables the examination of treatment effects through each
transition, on each event, and along each path. By analyzing data from the LEADER Trial, we
find that liraglutide significantly reduces the risk of cardiovascular and microvascular events. In the counterfactual world, there are |E| transitions that
are subject to intervention. Let a = (akg : (k, g) ∈ E) denote the levels of interventions, where
the entries of a may take different values.Where the paper says it
Where the paper says it
Figure S6 Summary
Figure S6 shows the estimated total treatment effects on nine individual events, with 95% confidence intervals. The direct effect on cardiovascular death is significant with a P-value of 0. p. 5 O D
E
M
O D
E
M
O D
E
M
O D
E
M
O D
E
M
q1 = O→D q2 = O→E→D
q3 = O→M→D q4 = O→E→M→D q5 = O→M→E→D
Figure 1: Five potential paths from baseline (O) to death (D) with expanded major adverse
cardiovascular events (E) and microvascular events (M) as two intermediate states. from initial states to state g as Qg. The effects operating through each pathway provide insight into multiple
questions, such as how the drug is effective and whether additional post-treatment interventions
should be recommended. For example, to study the direct effect, researchers wish to isolate
the effect of treatment itself on the primary endpoint, independent of any modification through
intermediate events.Where the paper says it
Where the paper says it
Figure S7 Summary
Figure S7 shows the estimated direct treatment effects on nine individual events (first occurrence), with 95% confidence intervals. The differential cumulative incidence of cardiovascular death is significant with a P-value of 0. p. 7 (2)
The primary object of interest is the time to reach a specific state j ∈ G (typically a terminal event
such as death), defined as T
a
j = inf{t ∈ [0, t∗
] : Xa
(t) = j} if state j is on the trajectory X
a
(t
∗
),
otherwise we set T
a
j = ∞. The primary estimand is the counterfactual cumulative incidence
function (CIF) of state j ∈ G, given by
F
a
j
(t) := P(T
a
j ≤ t) = P(j ∈ X
a
(t)), t ∈ [0, t∗
]. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition.Where the paper says it
Where the paper says it
Table S3 Summary
Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values of hypothesis tests for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0. p. 3 There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. There are several advantages to evaluating treatment effects through hypothetical interventions. First, by intervening in a single transition and comparing changes in cumulative incidence functions, one can determine whether the treatment affects this transition. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death. We assume that each state is visited at most once on any path. For example,
Figure 1 shows the multi-state structure for our motivating data, where O denotes the initial state,
E denotes expanded major adverse cardiovascular events, M denotes microvascular events, and
D denotes death.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Table S3 Summary (Second Instance)
The second instance of Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0.2975. The differential cumulative incidence of cardiovascular death shows a P-value of 0.0289.
Table S3 Summary (Third Instance)
The third instance of Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0.2975. The differential cumulative incidence of cardiovascular death shows a P-value of 0.0289.
Table S3 Summary (Fourth Instance)
The fourth instance of Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0.2975. The differential cumulative incidence of cardiovascular death shows a P-value of 0.0289.
Table S3 Summary (Fifth Instance)
The fifth instance of Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0.2975. The differential cumulative incidence of cardiovascular death shows a P-value of 0.0289.
Table S3 Summary (Sixth Instance)
The sixth instance of Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0.2975. The differential cumulative incidence of cardiovascular death shows a P-value of 0.0289.
Table S3 Summary (Seventh Instance)
The seventh instance of Table S3 reports the first occurrences of individual events stratified by treatment groups, and P-values for the direct treatment effect on each event. The differential cumulative incidence of non-fatal myocardial infarction shows a P-value of 0.2975. The differential cumulative incidence of cardiovascular death shows a P-value of 0.0289.
Table S3 Summary (EIF vs Reg)
Table S1 compares the bias and standard deviation (SD) for the EIF-based estimator and plug-in estimator for F a j(t) at time points t = 3, 6, 9, 12. The SD EIF is 0.015. The SD Reg – × is 0.037.
Table S2 Summary (Total Effects)
Table S2 shows the estimated coefficients in four Cox models for non-fatal EMACE, MVE, direct CVD, and other-cause death. The total treatment effects on EMACE and MVE are significant. The total effect on death is insignificant.
Table S2 Summary (Path Effects)
Figure S3 shows the estimated path-specific effects on death by intervening in each path from the initial state to death. The effect along the path O→E→D is P = 0. p. 3 Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment. Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect. Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment. Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect. Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment. Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect. Second, by intervening in
the transitions into a state, one can identify which events are affected by the treatment. Third,
by intervening in all the transitions along a path, one can estimate the path-specific treatment
effect.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Table S3 Summary (Path Effects)
Figure S3 shows the estimated path-specific effects on death by intervening in each path from the initial state to death. The effect along the path O→E→D is P = 0.5197. The effect along the path O→E→M→D is P = 0.6272.
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
-
Transition hazard modeling for multi-state data can be improved by assuming that
the transition hazard obeys the extended Markov property,
which restricts it to depending only onunordered states in the path
rather than the full past transition times. This allows for more tractable semiparametric estimation of hazards. -
Estimators for pathway treatment effects can be made more robust by using an Efficient Influence Function (EIF) that is
multiply robust and semiparametrically efficient.
This estimatordoes not require modeling of treatment-induced confounding,
which is a key advantage over traditional mediation frameworks. -
The system can be extended to infer dynamic treatment policies under sequential ignorability, allowing it to determine the optimal decision at each stage of disease progression by comparing
counterfactual CIFs under all possible dynamic treatment policies.
-
The AI system can estimate path-specific effects by decomposing the total effect into
interventional effects arising from sequentially intervening at each transition,
enabling researchers to distinguish between direct, event-specific, and path-specific treatment impacts. -
Sensitivity analysis using frailty models can be incorporated to account for potential unmeasured confounding, which helps in determining if observed treatment effects are robust even when assuming
the randomized intervention is performed conditional on both observed covariates and frailty.
Abstract
During clinical trials evaluating a drug's effect on a survival endpoint, intermediate events often occur before the primary event. The treatment can affect the primary event through multiple pathways via intermediate events. With multiple intermediate events, each is a treatment-induced confounder of transitions out of the others, so we cannot identify natural or separable effects. We instead consider stochastic interventions that assign each transition the identifiable observed-history hazard of either arm. Contrasts between the resulting counterfactual cumulative incidences define transition-, event-, and path-targeted effects, identifiable without cross-world or isolation assumptions. Based on the efficient influence functions for the counterfactual cumulative incidences under hypothetical interventions, we construct multiply robust and semiparametrically efficient estimators for these targeted treatment effects. Our proposed framework quantifies treatment effects through each transition, on each event, and along each path. By analyzing data from the LEADER trial, we find that liraglutide significantly reduces the risk of cardiovascular and microvascular events. The reduction in all-cause mortality is primarily mediated by its effects on expanded major adverse cardiovascular events.
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