📊 statistics

Time Partitioning in Target Trial Emulation

This paper demonstrates that selecting an appropriate time partitioning in target trial emulation is critical to balancing model dimensionality and causal structure validity, offering practical guidance to avoid the pitfalls of using data resolution as a default rather than tailoring the granularity to the specific clinical context.

Original authors: Harold Tankpinou Zoumenou (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Paris, France), Simon Ferreira (Sorbonne Universite, Inserm, Institut Pierre-Louis d

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Harold Tankpinou Zoumenou (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Paris, France), Simon Ferreira (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Paris, France), Charles Assaad (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Paris, France), Nathanael Lapidus (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Departement de Sante Publique, Hopital Saint-Antoine, APHP, Paris, France), Daria Bystrova (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Paris, France), Benjamin Glemain (Sorbonne Universite, Inserm, Institut Pierre-Louis d'epidemiologie et de sante publique, Departement de Sante Publique, Hopital Saint-Antoine, APHP, Paris, France, Paris Brain Institute-ICM, Inserm, Inria, Sorbonne Universite, Paris, France)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: Does a new medicine actually save lives, or do the people who took it just happen to be healthier to begin with?

In the world of medical research, we can't always run a perfect experiment where we flip a coin to decide who gets the medicine and who doesn't (a Randomized Controlled Trial). So, researchers use "Target Trial Emulation." This is like building a virtual simulation of that perfect experiment using real-world data from hospitals.

The problem? Real-world data is messy and happens in real-time. To analyze it, researchers have to chop time into little chunks, like slicing a loaf of bread. This paper is about how big those slices should be.

Here is the story of the paper, broken down into simple concepts:

1. The Bread Slicing Problem (Time Partitioning)

Imagine you are watching a movie to see if a character survives a storm.

Too Fine (The Microscope): If you slice time into seconds, you have millions of slices. Your computer gets overwhelmed trying to calculate the odds for every single second. It's like trying to count every single grain of sand on a beach to estimate the size of the beach. It's too much work, and you might get lost in the details.
Too Coarse (The Telescope): If you slice time into years, you miss the action. Maybe the medicine worked in the first month, but the patient died in the second. If you only look at "Year 1," you can't tell what happened when. It's like watching a movie where the screen is black for 11 months and then shows one frame. You miss the cause-and-effect.

The Paper's Lesson: You need the Goldilocks slice size. Not too small, not too big. It depends on how fast the medicine works.

2. The Two Stories (The Clinical Scenarios)

The authors tested their theory with two very different "movies."

Story A: The Slow-Acting Cancer Drug

The Scenario: A drug for metastatic cancer. It takes three months to show any effect on a CT scan.
The Analogy: Imagine a gardener planting a tree. You water it today, but you won't see the leaves grow for three months.
The Mistake: If you slice time into days, you are asking, "Did the water I gave today make the tree grow today?" The answer is obviously no. But by slicing it daily, you create a massive, confusing spreadsheet with thousands of variables.
The Solution: Slice time into months. Since the drug takes months to work, a monthly slice makes sense. It simplifies the math without losing the truth.

Story B: The Emergency ECMO Machine

The Scenario: A life-support machine (ECMO) for patients with severe breathing failure. These patients can die within hours.
The Analogy: Imagine a firefighter putting out a fire. If they don't act right now, the house burns down in minutes.
The Danger: If you slice time into weeks, you miss the critical moment. The patient might have died on Tuesday, but the machine was started on Wednesday. If you look at the whole week, you might think the machine saved them, when actually, they were already gone.
The Twist: Here, the outcome (death) can happen before the treatment decision for the next day is even made. This creates a "causal loop" (a chicken-and-egg problem) that breaks standard math tools.

3. The "Cloning" Trick (CCW)

To fix these messy timelines, researchers often use a clever trick called Cloning-Censoring-Weighting (CCW).

How it works: Imagine you take every patient in the study and make two clones of them.
- Clone A is forced to take the medicine every day.
- Clone B is forced to never take the medicine.
The Rule: You watch both clones. If Clone A stops taking the medicine (or Clone B starts taking it), you "censor" (delete) that clone from the study because they broke the rules.
The Catch: This trick only works if the medicine doesn't kill or save the patient during the same time slice.
- In Story A (Slow Drug): The drug doesn't kill you today. So, the cloning trick works perfectly.
- In Story B (Emergency): The patient might die before the doctor can even decide to start the machine for the next day. If you use the cloning trick here, you get the wrong answer because the "death" happened before the "treatment decision" for that specific slice could be made. The math breaks.

4. The Big Takeaway

The authors are telling researchers: Don't just use the data you have because it's convenient.

If your data is recorded every second, don't automatically analyze it every second.
Ask yourself: "How fast does this treatment work?"
- If it's slow (like a cancer drug), group the time into bigger chunks (weeks or months) to make the math easier and clearer.
- If it's fast (like emergency surgery), you must keep the time slices small, but you have to be very careful about the math you use, because the "cloning trick" might fail.

Summary Metaphor

Think of time partitioning like zooming in on a photo.

If you zoom in too much (daily data), the image is pixelated and blurry with too much noise.
If you zoom out too far (yearly data), the subject disappears.
The goal is to find the perfect zoom level where you can clearly see the relationship between the medicine and the outcome, without the picture getting distorted or the computer crashing.

The Bottom Line: In medical research, the way you slice time changes the answer you get. Choose your slices wisely based on the biology of the disease, not just the format of the spreadsheet.

1. Problem Statement

Target trial emulation is a cornerstone of causal inference using observational data, allowing researchers to estimate treatment effects while addressing time-varying confounding and immortal time bias. However, a critical methodological gap exists regarding time partitioning (the discretization of time into intervals for analysis).

The Dilemma: Researchers often default to the highest available data resolution (e.g., daily data).
- Too Fine: High-resolution partitioning (e.g., daily) leads to excessive model dimensionality, requiring the estimation of a vast number of parameters, which can cause overfitting and statistical inefficiency.
- Too Coarse: Coarse partitioning (e.g., monthly or yearly) may obscure the causal structure. Specifically, if a treatment affects the outcome (or vice versa) within a single time period, the resulting causal graph may contain cycles or violate assumptions required for identification.
The Ambiguity: Current reporting guidelines do not explicitly require researchers to justify their time partitioning strategy or analyze its implications for the causal relationship between treatment and outcome within specific intervals. This is particularly problematic for methods like Cloning-Censoring-Weighting (CCW), where the validity of assumptions depends on whether treatment influences the outcome within the same time period.

2. Methodology

The authors utilize the framework of Structural Causal Models (SCMs) and Directed Acyclic Graphs (DAGs) to formalize the problem.

Formal Framework:
- The study defines time as a sequence of periods $t = 1, \dots, T$ .
- Variables include treatment ( $X_t$ ), outcome ( $Y_t$ , e.g., death), and confounders ( $C$ ).
- The authors use the do-operator to define causal queries (e.g., $P(Y|do(X))$) and derive estimands (expressions using only observed data) via do-calculus.
- They employ Ancestral Multi-World Networks (AMWN) to test conditional independence assumptions required for CCW validity.
Comparative Case Studies:
The paper analyzes two distinct clinical scenarios to illustrate how the direction of causality within a time period dictates the appropriate partitioning:
1. Scenario A (Metastatic Cancer): A routine treatment where the effect on mortality is delayed (visible only after 3 months). Here, treatment ( $X_t$ ) does not influence outcome ( $Y_t$ ) within the same day/period.
2. Scenario B (ECMO in ARDS): A critical care intervention where treatment can influence survival within hours. Here, treatment ( $X_t$ ) does influence outcome ( $Y_t$ ) within the same period, creating a potential causal cycle if periods are too long.
Simulation:
- The authors simulate data for both scenarios (1,000 datasets of 1,000 patients each).
- They compare the bias of CCW against a non-parametric maximum likelihood estimator derived directly from the SCM estimand.

3. Key Contributions

A. Formalizing the Impact of Time Partitioning on Causal Structure

The paper demonstrates that the choice of time granularity is not merely a statistical decision but a causal modeling decision.

Fine Partitioning: Reduces the risk of causal cycles within periods but increases dimensionality (number of parameters).
Coarse Partitioning: Reduces dimensionality but risks introducing causal cycles (e.g., $X_t \to Y_t$ and $Y_t \to X_t$ ) if the treatment and outcome influence each other within the interval. Cycles violate the standard SCM assumption of acyclicity, rendering standard identification impossible without extensions.

B. Validity Conditions for Cloning-Censoring-Weighting (CCW)

The study provides a rigorous graphical proof regarding the validity of CCW:

Valid: CCW is valid when treatment does not influence the outcome within the same time period ( $X_t \not\to Y_t$ ). In this case, the conditional exchangeability assumption holds.
Invalid: CCW is invalid when treatment influences the outcome within the same time period ( $X_t \to Y_t$ ). In this scenario, the outcome ( $Y_t$ ) acts as a collider that opens a backdoor path, violating the conditional independence required for CCW.

C. Practical Guidance on "Grace Periods" and Partitioning

The authors clarify how to handle "grace periods" (time allowed for treatment initiation) and how to aggregate time units:

In Scenario A (delayed effect), the authors show that aggregating daily data into 28-day periods drastically reduces the number of parameters (from ~10,000 to ~28) without introducing bias, provided the treatment effect does not manifest within the 28-day window.
In Scenario B (immediate effect), aggregating data creates a causal cycle ( $X_t \leftrightarrow Y_t$ ). The authors argue that in such cases, one must either keep periods very short (daily) or accept that standard CCW cannot be used.

4. Results

Dimensionality Reduction: In the cancer scenario, switching from daily to 28-day partitioning reduced the parameter count by a factor of ~378, making estimation feasible without sacrificing validity (since $X_t \not\to Y_t$ ).
CCW Bias in Scenario A (No within-period effect):
- Both CCW and the SCM-based estimator showed negligible bias.
- Estimated bias for CCW: -0.06% (95% CI: -0.24% to 0.14%).
- Conclusion: CCW is valid when $X_t \not\to Y_t$ .
CCW Bias in Scenario B (Within-period effect):
- The SCM-based estimator remained unbiased.
- CCW exhibited significant bias.
- Estimated bias for CCW: -7.94% (95% CI: -8.13% to -7.76%).
- Conclusion: CCW fails when $X_t \to Y_t$ within the period because the censoring mechanism induces selection bias that cannot be corrected by weighting.

5. Significance and Implications

Methodological Shift: The paper argues against the default practice of using the finest available data resolution. Instead, researchers should actively specify a time partitioning strategy based on the clinical plausibility of within-period causal effects.
Causal Graph Awareness: It emphasizes that trial emulation requires explicit causal graphs (DAGs) to determine if a chosen time interval introduces cycles. If a cycle exists (e.g., treatment affects death, and death prevents further treatment observation within the same interval), standard methods like CCW are invalid.
Recommendations for Researchers:
1. Analyze the Lag: Determine if the treatment effect is immediate or delayed.
2. Widen Periods if Safe: If the treatment does not affect the outcome within the period, widen the time intervals to reduce dimensionality.
3. Avoid CCW for Immediate Effects: If treatment influences the outcome within the period, do not use CCW; instead, use methods that respect the specific causal structure (e.g., g-computation or direct estimands derived from SCMs).
4. Report Explicitly: Future guidelines should mandate the description of time partitioning strategies and the justification of the causal assumptions (specifically regarding within-period causality) underlying the chosen method.

In summary, this paper bridges the gap between statistical efficiency and causal validity in target trial emulation, providing a rigorous framework for selecting time partitions that ensure both identifiable and unbiased treatment effect estimates.