Time Partitioning in Target Trial Emulation

arXiv:2603.15924 · stat.ME · Submitted 2026-03-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

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

Tom: 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 Partitioning in Target Trial Emulation".

Jane: Here's the gist of the paper. Time partitioning enables researchers to handle time-varying confounders and immortal time bias with appropriate methods <ref:2603.15924#pg2> The study aims to explore issues related to time partitioning and provide guidance for trial emulation <ref:2603.15924#pg2> Methodological Challenges of Time Discretization The paper investigates the challenges introduced by partitioning time in target trial emulation, showing that the choice of partition can be too fine or too coarse depending on the clinical context <ref:2603.15924#pg4> A first challenge is the increase in model dimensionality because temporal segmentation multiplies variables and parameters to be estimated <ref:2603.15924#pg5> Another challenge arises from grace periods, which can be accommodated by methods like cloning-censoring-weighting (CCW) <ref:2603.15924#pg8> The literature lacks explicit recommendations for describing the temporal partitioning strategy or analyzing its implications for the causal relationship between treatment and outcome within each time period <ref:2603.15924#pg9> Implications of Partitioning Strategy The paper demonstrates that different assumptions about the causal relationship within a time interval lead to distinct graphical structures and choices of temporal partitioning <ref:2603.15924#pg9> Some studies assume the outcome influences treatment within a time interval, whereas others assume treatment influences the outcome within that same interval <ref:2603.15924#pg9> The paper shows that cloning–censoring–weighting may not be valid when treatment influences outcome within study periods through simulations <ref:2603.15924#pg9> Guidance for Appropriate Partitioning The authors provide practical guidance for actively specifying an appropriate time partitioning rather than using the available data resolution as a default <ref:2603.15924#pg9> They recommend widening time periods as long as no causal cycle occurs within periods <ref:2603.15924#pg9> The study concludes that researchers should discuss the causal relationship between treatment and outcome within study periods and use appropriate estimation procedures according to this relationship <ref:2603.15924#pg9> Validity of Estimation Procedures The paper evaluates the validity of CCW in different scenarios, finding that it cannot be used when treatment influences outcome within a time period <ref:2603.15924#pg9> Simulations show that for the second clinical scenario where outcome does not influence treatment, CCW can still yield low bias <ref:2603.15924#pg9> The study highlights that widening time periods decreases model dimensionality but can hinder estimation if two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9> The study suggests that time partitioning should be carefully chosen rather than defaulting to the available data resolution <ref:2603.15924#pg9> In conclusion, the paper highlights that time partitioning should be carefully chosen when emulating a trial, rather than defaulting to the available data resolution, which may be too fine or too coarse <ref:2603.15924#pg9> Researchers should discuss the causal relationship between treatment and outcome within study periods, widen these periods as long as no causal cycle occurs within periods, and use appropriate estimation procedures according to the causal relationship between treatment and outcome within periods <ref:2603.15924#pg9> The paper also suggests that in the presence of time-dependent confounders, estimands should be adapted as described in works on sequential interventions and g-computation <ref:2603.15924#pg9> Pruning redundant reasoning tokens The study uses the framework of SCMs to explore issues related to time partitioning in target trial emulation <ref:2603.15924#pg2> It shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9> How the training signal is built The estimand for Q0 in the first clinical scenario is derived through several steps involving law of total probability and rules of do-calculus to remove do-operators <ref:2603.15924#pg10> The final estimand for Q0 in the first clinical scenario is given by Xc P(C = c) YT k=1 P(Yk = 0Yt The estimand for Q1 in the first clinical scenario is given by Xc P(C = c) YT k=1 P(Yk = 0Yt The estimand for the ATE is expressed as a difference between Q1 and Q0 <ref:2603.15924#pg10> Where the gains come from The simulation results showed that for both procedures, the confidence intervals for the bias contained zero, indicating low bias <ref:2603.15924#pg9> More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was-0.06% (95% confidence interval: -0.24% to 0.14%) for CCW and-0.06% (95% confidence interval: -0.24% to 0.13%) for the procedure directly using equation 3 <ref:2603.15924#pg9> In the second clinical scenario, the estimated bias was-7.94% (95% confidence interval: -8.13%) for CCW and 0.02% (95% confidence interval: -0.21% to 0.26%) for the procedure based on equation 5 <ref:2603.15924#pg9> The paper also notes that the absence of an arrow from Yt to Xt is more plausible in certain situations, such as when mortality without treatment is low <ref:2603.15924#pg9> The study found that the confidence interval for the bias of CCW did not contain zero in one scenario <ref:2603.15924#pg9> Pruning redundant reasoning tokens The paper shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9> How the training signal is built The paper shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9> Where the gains come from The simulation results showed that for both procedures, the confidence intervals for the bias contained zero, indicating low bias <ref:2603.15924#pg9> More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was-0.06% (95% confidence interval: -0.24% to 0.14%) for CCW and-0.06% (95% confidence interval: -0.24% to 0.13%) for the procedure directly using equation 3 <ref:2603.15924#pg9> The second scenario showed a bias of-7.94% for CCW and 0.02% for the estimator based on equation 5 <ref:2603.15924#pg9> The paper also notes that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9> Pruning redundant reasoning tokens The paper shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.

Title and authors: Tom: So, let’s talk about who wrote this—Harold Tankpinou Zoumenou and his colleagues. They are researchers from places like Sorbonne Université and Inserm in France <ref:2603.15924#pg1>.

Jane: It’s interesting to see the team pulling together across different institutions, trying to tackle these complex observational problems that usually leave time-dependent confounding behind.

Lu: The authors are using the structural causal models framework, which is popular for encoding those direct causal relationships between variables through directed acyclic graphs or DAGs <ref:2603.15924#pg3>.

Meng: I wonder what the actual data looks like they used to set up these scenarios, because if the setup isn't right, any partitioning strategy could fail fast.

Lalam: The paper sets up two specific clinical scenarios to illustrate the challenges associated with time partitioning <ref:2603.15924#pg3>.

The paper's summary: Tom: The summary boils down to this: because time is changing in these scenarios, we need a way to manage that change without making the model impossibly large or biased.

Jane: It’s exploring how the choice of partitioning strategy directly impacts the causal structure we can actually estimate, which is crucial for getting reliable answers about treatment effects.

Lu: They explore different assumptions about what happens within a time interval; some studies assume one variable influences the other within that period, and others assume the opposite <ref:2603.15924#pg9>.

Meng: If we assume something that doesn't match reality in the clinical setting, our entire estimation procedure might be fundamentally flawed when we try to deploy it.

Lalam: They specifically look at a method called cloning-censoring-weighting or CCW and show that it might not be valid when treatment influences the outcome within those study periods <ref:2603.15924#pg9>.

The paper's improvements: Tom: The authors offer some pretty concrete guidance here: instead of just using whatever data resolution we have available, they suggest actively specifying the time partitioning strategy.

Jane: So, what does that mean practically for someone trying to set up a trial emulation study? Are they telling us to be more aggressive in widening those time windows?

Lu: They recommend widening the time periods as long as there isn't some kind of causal cycle occurring within those periods <ref:2603.15924#pg9>.

Meng: That makes sense from a resource perspective; wider periods mean fewer variables to track, which sounds like a way to keep the model dimensionality down for production.

Lalam: They conclude that researchers should really discuss that specific causal relationship between treatment and outcome within each study period and then pick an estimation procedure that matches that relationship <ref:2603.15924#pg9>.

Conclusion: Tom: So, to wrap up, the main point from "Time Partitioning in Target Trial Emulation" is that time partitioning needs to be chosen carefully rather than just defaulting to whatever data resolution we have available.

Jane: It’s about being intentional with those time windows and understanding what causal relationship exists between treatment and outcome within those periods before you start estimating anything.

Lu: The paper shows that if you widen the time periods, you reduce model dimensionality, but it can still hinder estimation if two variables are influencing each other inside that same period <ref:2603.15924#pg9>.

Meng: It’s a trade-off, right? Less complexity versus potential loss of information if the internal dynamics are complicated.

Lalam: And they point out that in some cases, like when there is a causal cycle between treatment and outcome, the specific causal effect we are interested in just isn't identifiable using standard structural causal models <ref:2603.15924#pg9>.

Tom: It’s a lot to take away for anyone trying to build robust methods for observational data analysis.

Jane: Definitely. So next time you look at setting up an emulation study, remember that the structure of your time partitioning matters a lot for the final estimate you get from this paper, "Time Partitioning in Target Trial Emulation."

Harold Tankpinou Zoumenou, *Simon Ferreira, Charles Assaad, ­Nathana¨el Lapidus, Daria Bystrova, Benjamin Glemain

Sorbonne University · Inserm Institut Pierre-Louis d’epidemiologie et de santé publique · H’pital Saint-Antoine APHP · Paris Brain Institute ICM Inserm Inria

stat.ME

Submitted: 2026-03-16

Updated: 2026-11-02

Comments: 31 pages; Previously this version appeared as arXiv:2609.34576 which was submitted as a new work by accident

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

The gist: Time partitioning enables researchers to handle time-varying confounders and immortal time bias with appropriate methods <ref:2603.15924#pg2> The study aims to explore issues related to time…

Key concepts

Time Partitioning
Dividing a continuous study timeline into distinct segments. The choice of segment length is crucial; it can either be too detailed (too fine) or too broad (too coarse), which affects model complexity and accuracy in analyzing treatment effects.
Causal Relationship Assumption
The paper explores different assumptions about how treatment and outcome interact within a time interval. Some studies assume the outcome affects treatment, while others assume the opposite. These differing assumptions dictate the graphical structure of causal models and influence which estimation methods are valid.
Cloning-Censoring-Weighting (CCW)
CCW is a method used to handle grace periods within time partitions. However, simulations showed it cannot be used when treatment influences the outcome within a study period. Its validity depends entirely on whether the assumed causal structure holds true for that specific interval.

Terminology

Summary

Time partitioning enables researchers to handle time-varying confounders and immortal time bias with appropriate methods <ref:2603. p. 0

Where the paper says it

tankpinou@iplesp. fr March 18, 2026 Abstract In target trial emulation, time partitioning enables researchers to handle time-varying confounders and immortal time bias with appropriate methods.

15924#pg2> The study aims to explore issues related to time partitioning and provide guidance for trial emulation <ref:2603. p. 0
Where the paper says it

fr March 18, 2026 Abstract In target trial emulation, time partitioning enables researchers to handle time-varying confounders and immortal time bias with appropriate methods. Based on two clinical scenarios, this study aimed to explore issues related to time partitioning and to provide guidance for trial emulation.

15924#pg2>

Methodological Challenges of Time Discretization

The paper investigates the challenges introduced by partitioning time in target trial emulation, showing that the choice of partition can be too fine or too coarse depending on the clinical context <ref:2603.15924#pg4> A first challenge is the increase in model dimensionality because temporal segmentation multiplies variables and parameters to be estimated <ref:2603. p. 1

Where the paper says it

Nevertheless, this partitioning of time introduces additional methodological challenges. A first challenge arises from the increase in model dimensionality, as temporal segmentation typically multiplies the number of variables and, consequently, the number of parameters to be estimated.

15924#pg5> Another challenge arises from grace periods, which can be accommodated by methods like cloning-censoring-weighting (CCW) <ref:2603. p. 1
Where the paper says it

To overcome this problem, the notion of grace period has been introduced. This concept can be accommodated within estimation methods such as cloning-censoring-weighting (CCW) (Cain et al.

15924#pg8> The literature lacks explicit recommendations for describing the temporal partitioning strategy or analyzing its implications for the causal relationship between treatment and outcome within each time period <ref:2603. p. 1
Where the paper says it

Current reporting guidelines for studies emulating randomized trials highlight the need to specify the identifiability assumptions of the causal effect of interest (Cashin et al. Nevertheless, they do not explicitly call for a description of the temporal partitioning strategy adopted, nor for an analysis of its implications for the causal relationship between treatment and outcome within each time period.

15924#pg9>

Implications of Partitioning Strategy

The paper demonstrates that different assumptions about the causal relationship within a time interval lead to distinct graphical structures and choices of temporal partitioning <ref:2603.15924#pg9> Some studies assume the outcome influences treatment within a time interval, whereas others assume treatment influences the outcome within that same interval <ref:2603. p. 1

Where the paper says it

Clarifying this issue is crucial, as the literature does not adopt a uniform position. Some studies assume the outcome to influence treatment within a time interval (Robins, Hern´an, and Brumback 2000; Hern´an, Brumback, and Robins 2000), whereas others assume that treatment influences the outcome within that same interval (Wanis et al.

15924#pg9> The paper shows that cloning–censoring–weighting may not be valid when treatment influences outcome within study periods through simulations <ref:2603. p. 0
Where the paper says it

When the partitioning is too coarse, the resulting causal structure may hinder effect estimation. We also show that cloning–censoring–weighting may not be valid when treatment influences outcome within study periods, and we confirm this through simulations.

15924#pg9>

Guidance for Appropriate Partitioning

The authors provide practical guidance for actively specifying an appropriate time partitioning rather than using the available data resolution as a default <ref:2603.15924#pg9> They recommend widening time periods as long as no causal cycle occurs within periods <ref:2603. p. 1

Where the paper says it

Nevertheless, they do not explicitly call for a description of the temporal partitioning strategy adopted, nor for an analysis of its implications for the causal relationship between treatment and outcome within each time period. More generally, these guidelines do not explicitly recommend the inclusion of a causal graph as part of the reporting framework.

15924#pg9> The study concludes that researchers should discuss the causal relationship between treatment and outcome within study periods and use appropriate estimation procedures according to this relationship <ref:2603. p. 0
Where the paper says it

When the partitioning is too coarse, the resulting causal structure may hinder effect estimation. We also show that cloning–censoring–weighting may not be valid when treatment influences outcome within study periods, and we confirm this through simulations.

15924#pg9>

Validity of Estimation Procedures

The paper evaluates the validity of CCW in different scenarios, finding that it cannot be used when treatment influences outcome within a time period <ref:2603.15924#pg9> Simulations show that for the second clinical scenario where outcome does not influence treatment, CCW can still yield low bias <ref:2603. p. 9

Where the paper says it

Similarly, one can set up AMWNs for all i, k ∈ {1, 2, 3}, and confirm that assumption 4 still holds (one can also show that assumption 4 holds in the original scenario of Figure 1). X1 X2 X3 Y1 Y2 Y3 Y x 2 Y x 3 (a) First clinical scenario (Xt ̸→ Yt) X1 X2 X3 Y1 Y2 Y3 Y x 1 Y x 2 Y x 3 (b) Second clinical scenario (Yt ̸→ Xt) Figure 2: Ancestral Multi-World Networks (AMWN) for testing conditional independence 4 in simplified versions of the clinical scenarios.

15924#pg9> The study highlights that widening time periods decreases model dimensionality but can hinder estimation if two variables influence each other within the same period <ref:2603. p. 0
Where the paper says it

When the partitioning is too fine, the dimensionality of the model is unnecessarily high. When the partitioning is too coarse, the resulting causal structure may hinder effect estimation.

15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603. p. 2
Where the paper says it

Therefore, the observed data can be used to estimate the causal effect whenever the causal query can be expressed as an estimand and the causal effect is called “identifiable”. Throughout this article, the letter X denotes the exposure of interest (treatment) and the letter Y denotes the outcome (death).

p. 2
Where the paper says it

Therefore, the observed data can be used to estimate the causal effect whenever the causal query can be expressed as an estimand and the causal effect is called “identifiable”. Throughout this article, the letter X denotes the exposure of interest (treatment) and the letter Y denotes the outcome (death).

p. 2
Where the paper says it

Therefore, the observed data can be used to estimate the causal effect whenever the causal query can be expressed as an estimand and the causal effect is called “identifiable”. Throughout this article, the letter X denotes the exposure of interest (treatment) and the letter Y denotes the outcome (death).

p. 2
Where the paper says it

Therefore, the observed data can be used to estimate the causal effect whenever the causal query can be expressed as an estimand and the causal effect is called “identifiable”. Throughout this article, the letter X denotes the exposure of interest (treatment) and the letter Y denotes the outcome (death).

p. 2
Where the paper says it

Therefore, the observed data can be used to estimate the causal effect whenever the causal query can be expressed as an estimand and the causal effect is called “identifiable”. Throughout this article, the letter X denotes the exposure of interest (treatment) and the letter Y denotes the outcome (death).

15924#pg9> The study suggests that time partitioning should be carefully chosen rather than defaulting to the available data resolution <ref:2603. p. 0
Where the paper says it

We also show that cloning–censoring–weighting may not be valid when treatment influences outcome within study periods, and we confirm this through simulations. In conclusion, we provide practical guidance for actively specifying an appropriate time partitioning in trial emulation, rather than using the available data resolution as a default.

15924#pg9>

Conclusion on Time Partitioning

In conclusion, the paper highlights that time partitioning should be carefully chosen when emulating a trial, rather than defaulting to the available data resolution, which may be too fine or too coarse <ref:2603.15924#pg9> Researchers should discuss the causal relationship between treatment and outcome within study periods, widen these periods as long as no causal cycle occurs within periods, and use appropriate estimation procedures according to the causal relationship between treatment and outcome within periods <ref:2603. p. 0

Where the paper says it

When the partitioning is too coarse, the resulting causal structure may hinder effect estimation. We also show that cloning–censoring–weighting may not be valid when treatment influences outcome within study periods, and we confirm this through simulations.

15924#pg9> The paper also suggests that in the presence of time-dependent confounders, estimands should be adapted as described in works on sequential interventions and g-computation <ref:2603.15924#pg9>

Pruning redundant reasoning tokens

The study uses the framework of SCMs to explore issues related to time partitioning in target trial emulation <ref:2603.15924#pg2> It shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603. p. 0

Where the paper says it

When the partitioning is too fine, the dimensionality of the model is unnecessarily high. When the partitioning is too coarse, the resulting causal structure may hinder effect estimation.

15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9>

How the training signal is built

The estimand for Q0 in the first clinical scenario is derived through several steps involving law of total probability and rules of do-calculus to remove do-operators <ref:2603.15924#pg10> The final estimand for Q0 in the first clinical scenario is given by Xc P(C = c) YT k=1 P(Yk = 0Yt The estimand for Q1 in the first clinical scenario is given by Xc P(C = c) YT k=1 P(Yk = 0Yt The estimand for the ATE is expressed as a difference between Q1 and Q0 <ref:2603. p. 6

Where the paper says it

Here, we focus on the ATE expressed as a difference, which completes the construction of our query: ATE = Q1 − Q0. It remains to remove the do-operators from the expressions for Q0 and Q1 in order to obtain an estimand.

15924#pg10>

Where the gains come from

The simulation results showed that for both procedures, the confidence intervals for the bias contained zero, indicating low bias <ref:2603.15924#pg9> More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was-0. p. 9

Where the paper says it

Over 1, 000 simulated datasets of 1, 000 patients each, we found that the confidence intervals for the bias of both procedures contained zero, indicating low bias (see Figure 3a). More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0.

p. 9
Where the paper says it

Over 1, 000 simulated datasets of 1, 000 patients each, we found that the confidence intervals for the bias of both procedures contained zero, indicating low bias (see Figure 3a). More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0.

06% (95% confidence interval: -0. p. 9
Where the paper says it

More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0. 06% (95% confidence interval: -0.

p. 9
Where the paper says it

More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0. 06% (95% confidence interval: -0.

p. 9
Where the paper says it

More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0. 06% (95% confidence interval: -0.

p. 9
Where the paper says it

More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0. 06% (95% confidence interval: -0.

24% to 0.14%) for CCW and-0. p. 1
Where the paper says it

To overcome this problem, the notion of grace period has been introduced. This concept can be accommodated within estimation methods such as cloning-censoring-weighting (CCW) (Cain et al.

p. 1
Where the paper says it

To overcome this problem, the notion of grace period has been introduced. This concept can be accommodated within estimation methods such as cloning-censoring-weighting (CCW) (Cain et al.

06% (95% confidence interval: -0.24% to 0.13%) for the procedure directly using equation 3 <ref:2603. p. 0
Where the paper says it

Keywords: Causal inference, Target trial emulation, Structural causal models, Clinical epidemiology 1 Introduction Target trial emulation now plays a central role in estimating treatment effects using observational data (Hern´an and Robins 2016). Representing treatment, confounders, 1 arXiv:2603.

p. 0
Where the paper says it

Keywords: Causal inference, Target trial emulation, Structural causal models, Clinical epidemiology 1 Introduction Target trial emulation now plays a central role in estimating treatment effects using observational data (Hern´an and Robins 2016). Representing treatment, confounders, 1 arXiv:2603.

15924#pg9> In the second clinical scenario, the estimated bias was-7. p. 9
Where the paper says it

In red, we highlight an open path between the variables of interest in the second clinical scenario. Finally, we simulated data compatible with this simplified example and estimated the bias of a nonparametric version of CCW (the weights described in Cain et al.

94% (95% confidence interval: -8. p. 9
Where the paper says it

More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0. 06% (95% confidence interval: -0.

13%) for CCW and 0. p. 1
Where the paper says it

To overcome this problem, the notion of grace period has been introduced. This concept can be accommodated within estimation methods such as cloning-censoring-weighting (CCW) (Cain et al.

02% (95% confidence interval: -0. p. 9
Where the paper says it

More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was -0. 06% (95% confidence interval: -0.

21% to 0.26%) for the procedure based on equation 5 <ref:2603. p. 0
Where the paper says it

Keywords: Causal inference, Target trial emulation, Structural causal models, Clinical epidemiology 1 Introduction Target trial emulation now plays a central role in estimating treatment effects using observational data (Hern´an and Robins 2016). Representing treatment, confounders, 1 arXiv:2603.

15924#pg9> The paper also notes that the absence of an arrow from Yt to Xt is more plausible in certain situations, such as when mortality without treatment is low <ref:2603.15924#pg9> The study found that the confidence interval for the bias of CCW did not contain zero in one scenario <ref:2603. p. 9
Where the paper says it

The details of the data-generating process and estimation procedures are provided in Supplement 3. Over 1, 000 simulated datasets of 1, 000 patients each, we found that the confidence intervals for the bias of both procedures contained zero, indicating low bias (see Figure 3a).

15924#pg9>

The paper shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9>

The paper shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.15924#pg9>

The simulation results showed that for both procedures, the confidence intervals for the bias contained zero, indicating low bias <ref:2603.15924#pg9> More precisely, with a true expected increase in survival of 23 % at the end of period 3 in the treatment group, the estimated bias was-0.06% (95% confidence interval: -0.24% to 0.14%) for CCW and-0.06% (95% confidence interval: -0.24% to 0.13%) for the procedure directly using equation 3 <ref:2603.15924#pg9> The second scenario showed a bias of-7. p. 9

Where the paper says it

In red, we highlight an open path between the variables of interest in the second clinical scenario. Finally, we simulated data compatible with this simplified example and estimated the bias of a nonparametric version of CCW (the weights described in Cain et al.

94% for CCW and 0. p. 1
Where the paper says it

To overcome this problem, the notion of grace period has been introduced. This concept can be accommodated within estimation methods such as cloning-censoring-weighting (CCW) (Cain et al.

02% for the estimator based on equation 5 <ref:2603. p. 9
Where the paper says it

Finally, we simulated data compatible with this simplified example and estimated the bias of a nonparametric version of CCW (the weights described in Cain et al. 2010 and the death hazards were estimated with saturated logistic regressions) and of a nonparametric maximum likelihood estimator of the ATE as expressed in equation 3 adapted to the 3-period simplified example (with each term estimated with a saturated logistic regression).

15924#pg9> The paper also notes that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603. p. 2
Where the paper says it

Therefore, the observed data can be used to estimate the causal effect whenever the causal query can be expressed as an estimand and the causal effect is called “identifiable”. Throughout this article, the letter X denotes the exposure of interest (treatment) and the letter Y denotes the outcome (death).

15924#pg9>

The paper shows that widening time periods decreased model dimensionality, but could also hinder estimation when two variables influence each other within the same period <ref:2603.15924#pg9> The authors also note that in the case of a causal cycle between treatment and outcome, the causal effect of interest is not identifiable in standard SCM frameworks <ref:2603.

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

Improvements for AI systems

  1. Time Partitioning Strategy Selection: The improved AI system can dynamically select an appropriate time partitioning strategy for target trial emulation based on whether treatment influences outcome within a period or vice versa, moving beyond default data resolution.

  2. Causal Graph Interpretation for CCW Validity: The system can use ancestral multi-world networks to test the validity of the cloning-censoring-weighting assumption (Assumption 4), specifically checking if "Yx i =Xk C, Xt<=k−1 = 0, Yt<=k = 0 for all i, k ∈ one... T" holds in complex scenarios.

  3. Estimand Identification and Estimation: The AI can compute the ATE using the derived estimands (equations 4 or 5) which correctly account for time-dependent confounding and the specific causal structures (Xt not -> Yt or Yt not -> Xt), leading to more accurate effect estimation than standard methods.

  4. Handling Causal Cycles: When a potential instantaneous cycle between treatment and outcome exists, the system can identify that the causal effect of interest is not identifiable within standard SCM frameworks, preventing flawed estimates based on cyclic models.

Abstract

In target trial emulation, the treatment strategies that patients follow are inferred from the treatments they actually receive in routine care. However, in most settings, the outcome may preclude the observation of planned treatment, giving rise to immortal time bias through misclassification of treatment strategy, while markers of treatment response may influence subsequent treatment decisions, giving rise to time-varying confounding. A key step toward unbiased treatment effect estimation is to partition follow-up into sufficiently short time intervals to unfold the feedback relationships involving treatment and represent the resulting causal relations with a directed acyclic graph. In this study, we present the possible within-interval causal orderings induced by this partitioning, discuss their causal implications, and assess their plausibility across clinical settings. For each causal ordering, we derive the corresponding g-formula. Using ancestral multi-world networks and simulations, we show that the standard cloning-censoring-weighting estimator is invalid when treatment affects the outcome within a time interval, and we propose a modified version of the method that restores its validity in this setting. Finally, we analyze the consequences of choosing time intervals that are either excessively wide or excessively narrow, thereby formally establishing the need for time partitioning and providing practical guidance for selecting an appropriate partition based on the clinical setting.

Sources

Related papers