Partitioning Time in Target Trial Emulation

arXiv:2609.34576 · stat.ME · Submitted 2026-09-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

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

Tom: We're looking at the paper: "Partitioning Time in Target Trial Emulation".

Jane: Here's the gist of the paper. In target trial emulation, partitioning follow-up into sufficiently short time intervals is a key step toward unbiased estimation of treatment effects <ref:2609.34576#pg2> because it unfolds the feedback relationships involving treatment and represents the resulting causal relations with a directed acyclic graph <ref:2609.34576#pg2>.

Tom: This study presents possible within-interval causal orderings induced by this partitioning, discusses their causal implications, and assesses their plausibility across clinical settings <ref:2609.34576#pg2>.

Jane: The Problem of Causal Biases Immortal time bias arises when treatment strategy is defined based only on treatment received prior to outcome occurrence, leading to misclassification of treatment strategy <ref:2609.34576#pg2>.

Tom: Time-varying confounding may also arise as markers of treatment response influence subsequent treatment decisions <ref:2609.34576#pg2>.

Paper summary: Tom: So, we’re looking at this paper by Tankpinou and colleagues called "Partitioning Time in Target Trial Emulation." Basically, they’re saying that in target trial emulation, which is how we try to estimate treatment effects using real-world data instead of a strict randomized trial setup, the way we define the treatment strategy based on what people actually get can cause problems.

Jane: Right. So it claims that if you only look at what patients *actually* receive in routine care to figure out their strategy, you might run into immortal time bias because sometimes the outcome happens before they even get the intended treatment, so you misclassify the treatment strategy Tom.

Lu: And it also points out that markers of how well a patient responds to treatment might influence what doctors do next, which creates something called time-varying confounding Jane. To fix this kind of feedback loop, they suggest breaking up the follow-up time into smaller intervals so you can map out the causal relationships using a directed acyclic graph Tom.

Meng: If we have to break up the time, what exactly are we looking at? Is this just about making sure we don't get stuck in a loop where treatment affects death, or maybe something else? Lu They introduce two specific ways the treatment and outcome can be ordered within those small time intervals.

Jane: That’s what they call within-interval causal orderings, specifically one where the outcome happens before the treatment starts within that interval, and another where the treatment happens before the outcome in that same small window Tom.

Tom: So it’s about which direction you assume the cause and effect go within those tiny slices of time, even if you don't know for sure Jane. They explore how these different orderings affect things like g-computation and cloning-censoring-weighting, which are the standard ways we try to estimate these effects Lu.

Lalam: I see that they use ancestral multi-world networks to look at these causal structures, and the main finding is that the standard cloning-censoring-weighting estimator isn't valid when treatment affects the outcome inside one of those time intervals Tom.

Jane: That’s a big deal because it means if you use the normal method, your results could be totally off if you don't account for this feedback loop Lu. They propose a modified version of that method that actually works in this setting and restores its validity Tom.

Meng: From an engineering standpoint, I’m interested in the practical guidance they give. What happens if we pick intervals that are way too wide or way too narrow? Does it just make the calculation take forever, or does it break the model entirely? Lu They show that excessively wide intervals can cause causal cycles between variables, which makes the treatment effect non-identifiable Tom.

Jane: And they also found that if the intervals are too short, you get a huge increase in variance because there’s an exponential increase in all those possible histories you have to consider Meng. So it sounds like we need a sweet spot for how long those time slices should be.

Tom: So the main conclusion here is that we really do need this time partitioning step when doing target trial emulation, and it gives us practical rules about picking the right interval length depending on what’s happening in the clinic Lu.

Jane: And they conclude that reporting this underlying causal structure, or DAG, needs to become a standard part of how we report these studies so people know what assumptions you made Tom.

Lu: It really forces us to be explicit about those assumptions when we try to estimate causal effects in these observational settings Jane. They also stress that the correct estimation procedure depends on which within-interval causal ordering you assume is true Meng.

Lalam: The paper says the appropriate estimation method isn't just one thing; it has to match the assumed structure, and that should be stated clearly in any reporting guidelines Tom.

Tom: So to wrap up, this paper on "Partitioning Time in Target Trial Emulation" is telling us that careful time partitioning is a necessary step to avoid major causal biases like immortal time bias and time-varying confounding Jane.

Jane: It gives us concrete advice on choosing the right interval size and which estimation method to use based on what we assume about the causal order within that small window Lu.

Meng: It sounds like it’s a very detailed roadmap for anyone trying to actually ship these kinds of models in a real-world setting, not just theoretical work Tom.

Lalam: I think the implication for people outside of pure statistics is that when we design these emulation studies, we need to be very specific about our time slices and our causal assumptions Jane.

Tom: Yeah, so it’s about moving from a general idea of "fix the bias" to a specific plan for how long those time intervals should be Lu.

Conclusion: Tom: So, we’re looking at this paper from Tankpinou and colleagues called "Partitioning Time in Target Trial Emulation."

Jane: It basically talks about how breaking up follow-up time into smaller chunks is really key to getting an unbiased estimate of treatment effects.

Tom: Yeah, they’re showing that without it, you run into these weird biases like immortal time bias and time-varying confounding.

Jane: They introduce these within-interval causal orderings, which means they assume different ways the treatment and outcome can be ordered within those small slices of time.

Tom: The main result is that the standard cloning-censoring-weighting estimator doesn't work when treatment affects the outcome inside one of those time intervals.

Jane: They propose a modified version of that method that actually fixes it, depending on which causal ordering you assume is true within those intervals.

Tom: And they give us some solid practical advice on choosing those time intervals—how wide or narrow they should be and what estimation procedure to use.

Jane: It also highlights the trade-off: too wide an interval can cause causal cycles, making the treatment effect unidentifiable, but too short increases the variance a lot because there are way more possibilities for covariate histories.

Tom: So it’s not just about finding *a* method that works; it’s about knowing how to choose the right time scale for your data before you even start estimating anything.

Jane: Exactly, and this is going to change how we think about setting up these real-world emulation studies in practice.

Harold Tankpinou Zoumenou, *, Simon Ferreira, Charles Assaad, David Hajage, Alexandra Beurton, Fabrice Carrat, Äthanaęel Lapidus, Ória Bystrova

Sorbonne University Institute of Medical Research Sorbonne Public Health Institute Equipe PEPITES Assistance Publique Hopitaux de Paris Sorbonne University Inria CNRS Paris Brain Institute

stat.ME

Submitted: 2026-09-28

Updated: 2026-11-02

Comments: This work was intended as a replacement of arXiv:2603.15924 and any subsequent updates will appear there

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

Importance score: 98/100

The gist: In target trial emulation, partitioning follow-up into sufficiently short time intervals is a key step toward unbiased estimation of treatment effects <ref:2609.34576#pg2> because it unfolds the…

Key concepts

Immortal Time Bias
This bias occurs when you define a patient's treatment strategy based only on the treatment they received *before* an outcome happens. If the outcome happens while they are still being treated, this method misclassifies what treatment strategy they actually followed, leading to incorrect results.
Time-Varying Confounding
This bias arises when factors that influence future treatment decisions change over time. For example, a patient's response to initial treatment might cause them to receive a different subsequent treatment. This makes it hard to isolate the true effect of the initial treatment.
Time Partitioning
Breaking down the total follow-up time into several smaller, distinct intervals is necessary to properly model complex feedback loops. This partitioning helps represent causal relationships using a directed acyclic graph, which is essential for unbiased analysis in target trial emulation.

Summary

In target trial emulation, partitioning follow-up into sufficiently short time intervals is a key step toward unbiased estimation of treatment effects <ref:2609. p. 0

Where the paper says it

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.

34576#pg2> because it unfolds the feedback relationships involving treatment and represents the resulting causal relations with a directed acyclic graph <ref:2609. p. 0
Where the paper says it

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.

34576#pg2>. This study presents possible within-interval causal orderings induced by this partitioning, discusses their causal implications, and assesses their plausibility across clinical settings <ref:2609. p. 0
Where the paper says it

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.

34576#pg2>.

The Problem of Causal Biases

Immortal time bias arises when treatment strategy is defined based only on treatment received prior to outcome occurrence, leading to misclassification of treatment strategy <ref:2609.34576#pg2>. Time-varying confounding may also arise as markers of treatment response influence subsequent treatment decisions <ref:2609. p. 1

Where the paper says it

More generally, immortal time bias may arise for any non-fatal outcome when treatment strategy is defined based only on treatment received prior to outcome occurrence. Another form of causal feedback may arise between treatment and markers of treatment response, such as biological monitoring measures, as these markers influence subsequent treatment decisions, leading to time-dependent confounding.

34576#pg2>. To avoid these issues, breaking apparent causal cycles between treatment and death or between treatment and time-varying confounders requires partitioning follow-up time into smaller intervals <ref:2609. p. 1
Where the paper says it

In such settings, the observed treatment strategy cannot be influenced by subsequent death or markers of treatment response. In all other cases, breaking the apparent causal cycles between treatment and death or between treatment and time-varying confounders requires partitioning follow-up time into smaller intervals, a necessary step toward unbiased estimation of treatment effects.

34576#pg2>.

Within-Interval Causal Orderings

The study introduces two reference within-interval causal orderings: (Zt, Yt, Xt) where the outcome precedes treatment within time intervals <ref:2609.34576#pg2> and (Zt, Xt, Yt) where treatment precedes the outcome within time intervals <ref:2609. p. 2

Where the paper says it

In Section 4, we evaluate the validity of nonparametric g-computation and cloning-censoring-weighting (CCW) under the two reference causal orderings (Robins 1986; Cain et al. Using ancestral multi-world networks (AMWNs) and simulations (Correa and Bareinboim 2025), we show that CCW may be biased when treatment precedes outcome within time intervals and propose a modification of the method to address this bias.

34576#pg2>. These orderings differ only in the relative position of treatment and outcome, with Zt time-varying confounders preceding both treatment Xt and outcome Yt for any time interval t <ref:2609. p. 3
Where the paper says it

Z1 additionally includes baseline confounders. The two orderings differ only in the relative position of treatment and outcome, with Xt occurring either before or after Yt.

34576#pg2>.

Estimation Procedures and Validity

The paper evaluates the validity of nonparametric g-computation and cloning-censoring-weighting (CCW) under these two reference causal orderings <ref:2609.34576#pg2>. Simulations show that the standard CCW estimator is invalid when treatment affects the outcome within a time interval, and a modified version of the method restores its validity in this setting <ref:2609. p. 0

Where the paper says it

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, 1 arXiv:2609.

34576#pg2>. The choice between standard and modified CCW depends on the assumed within-interval causal ordering <ref:2609. p. 9
Where the paper says it

2010), nonparametric g-computation, and a nonparametric version of the modified CCW introduced in Section 4. The standard and modified CCW procedures were applied unchanged under both orderings, whereas the g-computation estimand was adapted to the assumed within-interval structure (Equation 4 for the ordering (Zt, Yt, Xt) and Equation 5 for the ordering (Zt, Xt, Yt)).

34576#pg2>.

Practical Guidance for Partitioning

The study formally establishes the need for time partitioning by examining consequences of choosing intervals that are either excessively wide or excessively narrow <ref:2609.34576#pg2>. Excessively wide intervals can result in causal cycles between variables, leading to non-identifiability of the treatment effect <ref:2609. p. 1

Where the paper says it

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. Keywords: Time partitioning, Time discretization, Target trial emulation, Cloningcensoring-weighting, G-computation, Summary causal graphs, Ancestral multi-world networks 1 Introduction In randomized controlled trials, randomization ensures comparability between treatment groups with respect to both measured and unmeasured confounders.

34576#pg2>. Conversely, excessively short time intervals lead to increased estimator variance due to an exponential increase in possible covariate histories <ref:2609.34576#pg2>.

Conclusion and Recommendations

The paper provides guidance for partitioning time when designing a target trial emulation, including the choice of time interval length, the assumed within-interval causal structure, and the estimation procedure <ref:2609.34576#pg2>. It concludes that reporting the underlying DAG should become a standard component of reporting guidelines for target trial emulation studies <ref:2609. p. 2

Where the paper says it

Clarifying these issues is important for target trial emulation. Indeed, reporting guidelines require that any causal assumptions needed to estimate causal effects be explicitly stated (Cashin et al.

34576#pg2>. The appropriate estimation procedure depends on the assumed causal structure, and this should be explicitly stated in reporting guidelines <ref:2609. p. 2
Where the paper says it

Clarifying these issues is important for target trial emulation. Indeed, reporting guidelines require that any causal assumptions needed to estimate causal effects be explicitly stated (Cashin et al.

34576#pg2>.

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

Improvements for AI systems

  1. Improved treatment strategy inference for target trial emulation by using modified cloning-censoring-weighting when treatment affects outcome within a time interval to restore validity when standard methods fail. This allows AI systems to estimate treatment effects accurately in complex clinical settings where patients may receive or miss treatments dynamically within a follow-up period.

  2. Automated selection of appropriate time partitioning intervals based on clinical setting by analyzing the consequences of choosing excessively wide or excessively narrow intervals, providing practical guidance for selecting an appropriate partition. This enables AI systems to choose data discretization that minimizes non-identifiability and estimator variance in target trial emulation studies.

  3. Detection and mitigation of causal cycles between treatment and outcome or treatment and time-varying confounders by using ancestral multi-world networks to formally establish the need for time partitioning, thereby preventing estimation bias from these cycles. This allows AI systems to identify problematic causal structures in observational data before attempting estimation, leading to more robust causal inference.

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