📊 statistics

Estimating treatment duration effects via clone-censor-weight: a breast cancer case study

This paper formalizes the assumptions and compares estimation methods within the cloning-censoring-weighting (CCW) framework to address immortal time bias in estimating treatment duration effects, demonstrating its application and limitations through a breast cancer case study comparing 2 versus 5 years of adjuvant tamoxifen.

Original authors: Charlotte Voinot, Noémie Simon-Tillaux, Emma Torrini, Stefan Michiels, Bernard Sebastien, Clément Berenfeld, Julie Josse

Published 2026-05-20

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Charlotte Voinot, Noémie Simon-Tillaux, Emma Torrini, Stefan Michiels, Bernard Sebastien, Clément Berenfeld, Julie Josse

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

The Big Picture: The "What If" Game

Imagine you are a doctor trying to figure out the best length of time for a patient to take a specific breast cancer medication (tamoxifen). You want to know: Is it better to take it for 2 years or 5 years?

Ideally, you would run a perfect experiment (a Randomized Clinical Trial) where you flip a coin at the start: one group gets told "Stop at 2 years," and the other gets told "Stop at 5 years." You then watch who stays healthy.

But in the real world, we often only have observational data. This is like looking at a photo album of patients who already took the drug. Some stopped early because of side effects; some stopped because they moved away; some took it for 5 years. The problem is, these people weren't assigned by a coin flip. The people who managed to stay on the drug for 5 years might have been healthier to begin with. If you just compare the two groups, you might falsely conclude that "taking it longer is better," when really, the longer-takers were just the "survivors" who were healthy enough to keep going.

This paper introduces a clever statistical trick called Clone-Censor-Weight (CCW) to fix this mess and answer the "What If" question using real-world data.

Step 1: The Magic of Cloning (The "Twin" Strategy)

In the real world, a patient can't be in two places at once. They either took the drug for 2 years or 5 years. But to compare the two strategies fairly, the researchers use a statistical magic trick: Cloning.

Imagine taking every single patient in the database and creating a digital twin for them.

Twin A is assigned the strategy: "Take the drug for exactly 2 years."
Twin B is assigned the strategy: "Take the drug for exactly 5 years."

Now, both twins start with the exact same medical history, the same age, and the same health status. They are identical at the beginning. This is like setting up a fair race where both runners start at the same line.

Step 2: The "Stop Sign" (Artificial Censoring)

Here is where the real-world data gets messy.

Twin A (assigned 2 years) might actually keep taking the drug for 3 years in the real data.
Twin B (assigned 5 years) might stop after 1 year because of side effects.

To make the comparison fair, the researchers have to stop the clock (censor) on the twins the moment they deviate from their assigned rule.

If Twin A keeps taking the drug past year 2, the researchers say, "Okay, Twin A, your experiment is over. We can't count your data anymore because you broke the rule."
If Twin B stops at year 1, the researchers say, "Twin B, your experiment is over."

This is called Artificial Censoring. It's like a referee blowing a whistle and removing a player from the game because they stepped out of bounds.

The Problem: This creates a new bias. If you just throw away the data for the twins who broke the rules, you might be left with only the "super healthy" twins who naturally followed the rules perfectly. This is called Immortal Time Bias. It's like saying, "Look, everyone who finished the marathon is healthy," while ignoring everyone who quit halfway through because they were sick.

Step 3: The "Weight" (Fixing the Balance)

To fix the bias caused by throwing people out of the game, the researchers use Weights.

Imagine you have a bag of marbles representing your patients.

If a patient was very likely to break the rules (e.g., they had a history of side effects) but didn't break them, they are a rare, valuable marble.
If a patient was very likely to break the rules and did break them, they are a common marble.

The Weighting step gives more "importance" (a heavier weight) to the patients who stayed in the game against the odds. It's like saying, "This patient stayed for 5 years even though they had a high risk of stopping. Let's count their experience as if it represents 10 other people who were just like them."

By adjusting the weights, the researchers reconstruct a "virtual population" where everyone had an equal chance of following either the 2-year or 5-year rule, effectively removing the bias.

Step 4: The "Time-Varying" Twist

The paper also tackles a harder version of the problem. Sometimes, a patient's health changes during the treatment.

Maybe a patient gets a new symptom at year 2.
This symptom might make them stop the drug and might make them more likely to get sick later.

This is a "moving target." The researchers show that if you don't account for these changing health factors (time-varying confounders), your weights might still be wrong. They tested different mathematical recipes (like G-formula and IPCW) to see which one handles these moving targets best.

The Results: What Did They Find?

The authors tested their methods using computer simulations (creating fake data where they knew the "true" answer) and then applied it to a real French breast cancer database.

Naive comparisons fail: If you just look at who took the drug longer without fixing the bias, you get the wrong answer.
Cloning helps: Creating the twins removes the initial "survivor" bias.
Weighting is crucial: You must weight the data to account for people who were forced out of the study (either by stopping the drug or by dropping out of the medical records).
The Best Recipe: In complex situations where health changes over time, the method that combined Cloning + Artificial Censoring + Inverse Probability Weighting (IPCW) worked best. It was the most reliable way to estimate the true effect of treatment duration.

The Bottom Line

This paper doesn't say "You should take the drug for 5 years." Instead, it provides a better toolkit for scientists to ask that question using real-world data.

Think of it like this:

Old way: Looking at a photo of a finished race and guessing who would have won if they had run a different distance. (Biased and unreliable).
New way (CCW): Creating digital twins, forcing them to run specific distances, stopping them if they wander off course, and then mathematically adjusting the score to account for who was most likely to wander off.

The authors conclude that while this method is powerful, it requires careful handling. If the math models used for the "weights" are slightly wrong, the results can still be skewed. Therefore, they recommend checking the results with different models to ensure the conclusion is robust.

In short: They built a sophisticated statistical machine to simulate a perfect clinical trial using messy, real-world data, allowing doctors to better understand how long breast cancer patients should stay on medication.

Technical Summary: Estimating Treatment Duration Effects via Clone-Censor-Weight: A Breast Cancer Case Study

Problem Statement
Randomized clinical trials (RCTs) are the gold standard for evaluating treatment effects but often lack transportability to routine care due to strict eligibility criteria and controlled settings. Observational data are essential for evaluating real-world treatment strategies, particularly those involving treatment duration. However, estimating the causal effects of static duration strategies (e.g., "stop at 2 years" vs. "stop at 5 years") in observational settings is fraught with methodological challenges.

The primary difficulties include:

Time-varying confounding: Covariates measured during follow-up may affect subsequent treatment decisions and survival outcomes while being influenced by prior treatment.
Immortal time bias: Naive comparisons based on realized treatment duration condition on post-baseline events. Individuals must survive and remain event-free long enough to receive a specific duration, creating a spurious association between longer duration and better survival.
Informative censoring: Both natural censoring (loss to follow-up) and artificial censoring (deviation from the assigned strategy) can be informative, depending on patient history.
Model complexity: Existing methods like the parametric G-formula, marginal structural models (MSMs), and structural nested models (SNMs) face challenges regarding model misspecification, weight instability, and computational burden, particularly when grace periods (relaxed adherence rules) are required.

Methodology
The authors propose and evaluate a framework based on the Cloning–Censoring–Weighting (CCW) approach to emulate a target trial comparing static treatment duration strategies. The methodology is developed in two settings: (1) confounding captured solely by baseline covariates, and (2) confounding involving time-varying covariates.

The core procedure involves three steps:

Cloning: Each individual in the observational dataset is cloned into multiple copies, one for each treatment strategy being compared (e.g., $d_0$ and $d_1$ ). All clones share the same observed baseline and follow-up data but are assigned different hypothetical strategies.
Artificial Censoring: Clones are artificially censored at the first time point where their observed treatment history deviates from their assigned strategy. To address clinical reality, the authors incorporate "grace periods" or relaxed adherence rules, defining an admissible set of trajectories ( $A_d$ ) rather than requiring exact adherence.
Inverse Probability of Censoring Weighting (IPCW): To correct for the informative nature of both artificial and natural censoring, clones are weighted by the inverse probability of remaining uncensored. The authors distinguish between:
- Artificial censoring weights: Based on the probability of remaining compatible with the assigned strategy.
- Natural censoring weights: Based on the probability of remaining under observation.
- The authors investigate modeling these processes separately (using pooled logistic regression for artificial censoring and piecewise exponential models for natural censoring) versus jointly.

The paper compares several estimators for the Restricted Mean Survival Time (RMST) contrast:

IPCW-Kaplan-Meier: Uses the calculated weights in a weighted survival analysis.
G-formula (Outcome Regression): Models the outcome directly on the cloned data, adjusting for covariates, without explicit weighting for censoring.
Augmented IPCW (AIPCW): A doubly robust estimator combining outcome regression and IPCW.

Key Contributions and Simulation Results
The authors conducted extensive numerical experiments to assess the performance of these estimators under varying degrees of confounding and informative natural censoring.

Immortal Time Bias: The simulations confirm that naive approaches (filtering by observed duration) and G-formula approaches applied only to adherent individuals fail to correct for immortal time bias, resulting in severe overestimation of treatment benefits.
Baseline Confounding (Scenario 1): When confounding is limited to baseline covariates, multiple methods (IPCW, G-formula, AIPCW) perform well if models are correctly specified. However, the G-formula showed stability even under strong confounding, though it remained sensitive to model misspecification (e.g., using a Weibull model when the truth was exponential).
Time-Varying Confounding (Scenario 2): This setting proved more challenging.
- IPCW Performance: The IPCW estimator, specifically when modeling artificial and natural censoring separately (pooled logistic for artificial, piecewise exponential for natural), provided the most consistent results across scenarios. It relies on observed covariate trajectories rather than simulating counterfactual ones.
- G-formula Limitations: In the presence of time-varying confounding, the G-formula exhibited substantial bias, particularly near positivity violations. This is attributed to the accumulation of prediction errors when reconstructing full counterfactual covariate trajectories over time.
- Model Misspecification: Modeling both censoring mechanisms with a single regression model (e.g., a single pooled logistic model) introduced bias in complex settings, suggesting that separating the mechanisms is preferable despite potential variance inflation.
- AIPCW: While theoretically doubly robust, the implementation of AIPCW in this specific cloning framework with time-varying covariates was noted as computationally demanding and lacking dedicated software implementation in the current literature.

Real-World Application
The methodology was applied to a large French nationwide prospective cohort of patients with localized breast cancer to evaluate the effect of tamoxifen duration. The study aimed to emulate a trial comparing stopping treatment after 2 years versus continuing for 5 years. The authors utilized the CCW framework to handle the complex longitudinal treatment histories and adherence patterns observed in routine care, where nearly 55% of patients discontinue early.

Significance and Claims
The paper claims to provide a methodological framework for evaluating treatment duration strategies in observational data, specifically addressing the gap between baseline-only and time-varying confounding settings.

Practical Guidance: The authors conclude that while many approaches exist for baseline confounding, IPCW-based approaches appear to offer the best compromise between implementation complexity and empirical performance when time-varying confounders are present. They highlight that IPCW methods are less sensitive to the accumulation of errors associated with simulating full counterfactual covariate trajectories (as required by the G-formula).
Caveats: The authors remain modest regarding the universality of their recommendations. They emphasize that:
- The assumption of "admissible consistency" (that relaxed adherence rules do not alter the causal estimand) is strong and difficult to verify.
- No single estimator is uniformly recommended; conclusions should rely on sensitivity analyses across plausible model specifications.
- Challenges remain regarding missing data, positivity violations, and the computational burden of nuisance model estimation in complex longitudinal settings.

The work underscores the necessity of careful modeling of both artificial and natural censoring mechanisms and suggests that future methodological developments are needed to address model uncertainty and complex data structures in causal inference for treatment duration.