📊 statistics

Considerations for Estimating Causal Effects of Informatively Timed Treatments

This paper addresses the methodological gap in epidemiological studies regarding informatively timed treatment sequences by formalizing the issue, demonstrating that ignoring variable waiting times between treatments leads to bias, and showing how g-methods can correct for this by treating waiting times as time-varying confounders to ensure valid causal inference.

Original authors: Arman Oganisian

Published 2026-01-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arman Oganisian

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 trying to figure out which of two different training regimens helps marathon runners finish a race faster. But there's a catch: the runners don't start their next training phase on a fixed schedule, like "every Monday at 9 AM." Instead, they start their next phase whenever they feel strong enough.

If a runner feels great and recovers quickly, they start the next phase sooner. If they are exhausted or sick, they wait longer.

This paper argues that how long a runner waits between phases isn't just a random number; it's a huge clue about how well they are doing. If you ignore when they started their next phase and only look at what they did, you will get the wrong answer about which training regimen is actually better.

Here is a breakdown of the paper's main ideas using simple analogies:

1. The Problem: "Informative Timing"

In many medical studies, doctors don't give treatments at fixed times. They wait until the patient is ready.

The Scenario: Think of a patient recovering from chemotherapy. They need to wait until their heart is strong enough (measured by "ejection fraction") before getting the next dose.
The Trap: Patients who recover fast (short wait time) are usually the ones who are naturally stronger and tolerate the medicine well. Patients who recover slowly (long wait time) might be sicker.
The Mistake: If a researcher just compares the "Fast Waiters" to the "Slow Waiters" without accounting for why they waited, they might think the treatment is working better or worse than it actually is. The "wait time" itself is acting like a hidden variable that confuses the results. The paper calls this "Informative Timing."

2. The Solution: Treating "Wait Time" as a Clue

The author says we shouldn't ignore the wait time. Instead, we should treat the time between treatments as a time-varying confounder.

The Analogy: Imagine you are judging a cooking contest. If a chef takes a long break between courses, it might be because they are struggling with the recipe (bad outcome) or because they are being extra careful (good outcome).
The Fix: To judge the chef fairly, you can't just look at the final dish. You have to look at the whole story: the ingredients they used, their skill level, and how long they took between steps. If you adjust your math to include that "wait time," you can see the true effect of the cooking technique.

3. The Tool: "G-Methods"

The paper suggests using a specific set of statistical tools called g-methods (like Inverse Probability Weighting).

How it works: Think of this as a "fairness filter." The computer looks at every patient and asks: "Given how sick this person was and how long they waited, how likely were they to get the treatment they actually got?"
The Result: It creates a "virtual world" where the wait times are balanced out. It effectively says, "Let's pretend the slow-recovering patients started their next treatment at the same time as the fast-recovering ones, just to see what would happen." This removes the bias caused by the timing.

4. The "Death" Complication

The paper also deals with a grim reality: some patients die before they can get their second treatment.

The Challenge: If you only look at patients who survived long enough to get the second treatment, you are ignoring the people who died early. This is a huge mistake because the people who died early might have been the ones who needed the treatment most (or least).
The Fix: The proposed method includes everyone. It counts the "wait time" for those who died as well as those who survived. It ensures that the people who dropped out (either by dying or leaving the study) don't skew the results.

5. Continuous vs. Discrete Time

The paper shows that you can solve this problem whether you look at time as a smooth river (continuous) or as a series of stepping stones (discrete days/weeks).

The Finding: It doesn't matter which mathematical "lens" you use. As long as you account for the wait time between treatments, you get the same correct answer. The paper provides code and examples showing that both approaches work.

The Bottom Line

The author's main message is simple but critical:

Timing matters: In studies where treatments happen at different times for different people, the time between treatments is a major factor.
Don't ignore it: If you ignore the wait time, your results will be biased (wrong).
Fix it easily: You can fix this using standard statistical tools (g-methods) by simply adding the "wait time" into your calculations as a variable that needs to be adjusted for.

What the paper does NOT say:

It does not claim that a specific drug is better than another.
It does not tell doctors exactly how to treat patients in a hospital.
It does not promise that this will work for every single disease in the world.

It strictly focuses on the math and logic of how to analyze data correctly when the timing of events is part of the story, not just a background detail.

Technical Summary: Considerations for Estimating Causal Effects of Informatively Timed Treatments

Problem Statement
Epidemiological studies frequently aim to estimate the causal effects of sequential treatment decisions on survival outcomes, particularly in chronic disease management. A critical, often overlooked issue in these settings is that treatment decisions do not occur at fixed, pre-specified intervals. Instead, the timing of subsequent treatments varies across subjects in ways that are "informative" of both future treatment decisions and potential outcomes.

The paper argues that standard causal inference methods, such as G-methods (e.g., Inverse Probability of Treatment Weighting, IPTW), which are designed for time-varying confounding, may yield biased estimates if they fail to account for this "informative timing." Specifically, the waiting times between successive treatment courses act as time-varying confounders. For instance, in pediatric acute myeloid leukemia (AML) trials, the time required for a patient to recover from toxicity before starting the next chemotherapy course is indicative of their underlying health status. Patients who recover faster (shorter waiting times) may be more likely to tolerate a specific treatment and have better survival prospects, creating a backdoor path between treatment assignment and survival that standard models may miss.

Methodology
The author formalizes the problem using a potential outcomes framework with $K$ treatment courses. The core innovation is treating the waiting times between treatments ( $W_k$ ) as time-varying confounders that must be adjusted for to achieve valid causal identification.

Formalizing Informative Timing:
The paper defines the waiting time $W_1$ (time from treatment 1 to treatment 2) and $W_2$ (time from treatment 2 to death) as components of the survival time $T = W_1 + W_2$ . The causal structure is represented via a Directed Acyclic Graph (DAG) where $W_1$ confounds the effect of the second treatment ( $A_2$ ) on the subsequent survival time ( $W_2$ ).
- Identification Assumptions: The approach relies on sequential exchangeability assumptions. Crucially, the second treatment decision ( $A_2$ ) must be conditionally independent of the potential waiting time to death given the history, which now explicitly includes the observed waiting time $W_1$ .
- Positivity: The method requires that for any history (including the specific waiting time $W_1$ ), there is a non-zero probability of receiving either treatment option.
Continuous-Time Approaches:
The paper derives a G-formula for the survival probability $P(T^{a_1, a_2} > \tau)$ that incorporates sub-density functions of waiting times to account for competing risks (death vs. subsequent treatment).
- Estimation: Two primary estimators are discussed:
  - G-computation: Requires modeling the density of waiting times and covariates, which is complex in continuous time with competing risks.
  - Hájek IPTW: A simpler inverse-probability weighting estimator. The weights are constructed to adjust for both treatment assignment and the informative timing. For the second treatment, the propensity score model must condition on the waiting time $W_1$ from the first course.
Handling Censoring:
The methodology is extended to scenarios where patients may be censored between treatments. This requires an additional assumption of non-informative censoring and the inclusion of censoring weights in the IPTW estimator. The estimator sums contributions from patients who die after one treatment, those who die after two, and those who are censored, ensuring the target population is properly represented.
Discrete-Time Equivalence:
The paper demonstrates that these concepts translate to a discrete-time framework. By partitioning the time window into intervals, the waiting time is represented by the interval index $S$ at which the second treatment occurs. The discrete-time identification assumptions mirror the continuous-time ones, requiring conditioning on the number of intervals elapsed since the last treatment. The paper shows that discrete-time G-methods (using pooled logistic regression or marginal structural models) yield equivalent results to the continuous-time approaches when the waiting time (or interval index) is included in the propensity score models.

Key Contributions and Results

Formalization of Informative Timing: The paper explicitly identifies waiting times between treatments as a specific type of time-varying confounder that creates bias if ignored.
Bias Demonstration: Using synthetic data mimicking the AAML1031 pediatric AML trial, the author illustrates that "Unadjusted IPTW" (which ignores $W_1$ ) and "Naive" estimators (which restrict analysis to those completing both courses) produce significantly biased point estimates and poor confidence interval coverage (e.g., coverage dropping to 6.4% for naive estimators).
Correction via Adjustment: The "Adjusted IPTW," which includes waiting times in the propensity score models, successfully eliminates bias and restores coverage to nominal levels (approx. 94-95%).
Discrete vs. Continuous: The paper establishes that the need to adjust for informative timing is agnostic to the choice of modeling framework (continuous vs. discrete). Both approaches, when correctly specified to include the timing variable, produce consistent estimates.
Implementation: The author provides implementation guidance and publicly available code for applying these G-methods with waiting time adjustments.

Significance and Claims
The paper concludes with two primary messages:

Validity of Inference: Considering the timing of treatment decisions is essential for valid causal inference in epidemiological studies involving sequential treatments. Ignoring the informative nature of waiting times leads to biased estimates of treatment effects.
Feasibility of Correction: Correcting for informative timing is achievable using standard G-methods (specifically IPTW and G-computation) by treating the waiting times between treatments as time-varying confounders.

The author emphasizes that while the literature on irregular visit processes exists, the specific application to causal inference of sequential treatments where timing is informative has been under-addressed. The work bridges this gap, showing that standard causal tools can be adapted to handle these complex, real-world trial protocols where patients progress to the next course "as soon as clinically acceptable" rather than at fixed intervals.