Time-smoothed inverse probability weighted estimation of effects of generalized time-varying treatment strategies on repeated outcomes truncated by death

Sean McGrath, Takuya Kawahara, Joshua Petimar, Sheryl L. Rifas‐Shiman, Iván Díaz, Jason P. Block, Jessica G. YoungarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces "time-smoothed" inverse probability weighted (IPW) estimators to handle the "messy" reality of longitudinal electronic health record (EHR) data. By leveraging repeated, sparse outcome measurements and applying a "treatment-death isolation" assumption, the authors significantly increase the statistical precision of causal effect estimates for time-varying treatments, even when outcomes are truncated by death.

What Problem Does This Paper Address?

Estimating causal effects from real-world data like EHRs is notoriously difficult due to three overlapping challenges:

  1. Complex Treatments: Treatment strategies often change over time and depend on a patient’s evolving health status (e.g., a doctor might pause a medication if a patient becomes pregnant).
  2. Sparse & Informative Data: Outcomes (like weight) aren't measured on a fixed schedule. They are measured "repeatedly, non-monotonically, and sparsely." Furthermore, the fact that a measurement exists at all might be related to the patient's health, creating "informative" missingness.
  3. Truncation by Death: If a patient dies, their outcome (e.g., weight change) becomes undefined. Standard "total effect" models break down because you cannot compare a survivor to someone who is deceased.

Existing methods often address these separately. This paper seeks a unified approach that uses all available data points to improve the "signal-to-noise" ratio while properly accounting for the "truncation" caused by death.

Methods

The authors build upon the generalized g-formula and Inverse Probability Weighting (IPW). Their core innovations are:

  • Time-Smoothing: Instead of only looking at data from a specific target month (e.g., month 24), they fit a model across all follow-up times. This "smooths" the estimates, allowing data from month 22 or 23 to inform the estimate for month 24, which drastically reduces variance.
  • Handling Death via "Story-Led" Causal Inference: They argue that for many clinical questions, the most meaningful target is the effect among those who would survive regardless of treatment. They utilize a treatment-death isolation assumption (assuming the treatment doesn't directly cause death in the short term) to identify a "conditional separable effect."
  • Stacked vs. Non-stacked Estimators: They propose two ways to implement smoothing. The "stacked" version is particularly powerful as it pools data across different subsets of survivors at different time points to maximize efficiency.

The methods are implemented in a new R package called smoothedIPW.

Datasets

The paper evaluates the proposed methods using two primary sources:

  1. Simulated Longitudinal Data: 1,000 datasets of 1,000 individuals each, simulating 24 months of follow-up with treatment-confounder feedback, sparse measurements, and a high (70%) death rate by the end of the study.
  2. The MedWeight Study (PCORnet): A real-world dataset of 14,882 individuals with depression. The goal was to compare the effects of two antidepressants (sertraline vs. citalopram) on weight change over 24 months. This dataset featured high missingness (up to 84% at month 24).

Key Insights and Ideas

  • Efficiency is King: In EHR data, the "target" time point often has very few actual measurements. By smoothing over the entire timeline, researchers can salvage information from the "off-target" months.
  • The "Aha" Moment on Death: The authors highlight that "total population effects" are often a mathematical fiction when death occurs. By focusing on a "story-led" approach, they align the math with the clinical reality: a doctor wants to know how a drug affects weight in patients who are still alive to be weighed.
  • Bias-Variance Tradeoff: Smoothing introduces a small risk of bias if the time-model is wrong, but the simulation shows that the massive gain in precision (narrower confidence intervals) usually outweighs this risk in sparse datasets.

Key Results

  • Massive Precision Gains: In simulations, the time-smoothed approaches were 33 to 50 times more efficient (lower variance) than standard IPW at later time points.
  • Real-World Impact: In the MedWeight analysis, the standard (non-smoothed) IPW produced a 95% confidence interval for weight change at 24 months that was 4.66 kg wide. The stacked, time-smoothed version narrowed that interval to just 0.50 kg.
  • Consistency: Despite the increased precision, the smoothed methods arrived at similar point estimates to traditional methods, suggesting the smoothing model didn't introduce significant bias in this application.

Limitations and Future Work

  • Assumption Sensitivity: The "treatment-death isolation" assumption is more plausible for short-term outcomes (e.g., 6 months) than long-term ones (24 months). If the treatment actually influences mortality, the estimates could be biased.
  • Model Misspecification: The smoothing relies on a parametric model for the outcome over time. If the true trajectory of the outcome is highly irregular, a simple linear or quadratic smoothing function might fail.
  • Unmeasured Confounding: Like all IPW methods, this assumes all factors influencing both treatment and outcome are measured. The authors suggest that incorporating past outcome measurements (even if sparse) is a critical next step to mitigate this.
  • Future Directions: The authors suggest exploring more flexible smoothing techniques, such as restricted cubic splines, and extending the framework to Doubly Robust estimators (like TMLE).

Summary updated

Abstract

Researchers are often interested in estimating effects of generalized time-varying treatment strategies on the mean of an outcome at one or more selected follow-up times of interest. For example, the Medications and Weight Gain in PCORnet (MedWeight) study aimed to estimate effects of adhering to flexible medication regimes on future weight change using electronic health records (EHR) data. This problem presents several methodological challenges that have not been jointly addressed in the prior literature. First, this setting involves treatment strategies that vary over time and depend dynamically and non-deterministically on measured confounder history. Second, the outcome is repeatedly, non-monotonically, informatively, and sparsely measured in the data source. Third, some individuals die during follow-up, rendering the outcome of interest undefined at the follow-up time of interest. In this article, we pose a range of inverse probability weighted (IPW) estimators targeting effects of generalized time-varying treatment strategies in truncation by death settings that allow time-smoothing for precision gain. We conducted simulation studies that confirm precision gains of the time-smoothed IPW approaches over more conventional IPW approaches that do not leverage the repeated outcome measurements. We illustrate an application of the IPW approaches to estimate comparative effects of adhering to flexible antidepressant medication strategies on future weight change. The methods are implemented in the accompanying R package, smoothedIPW.
BibTeX
@misc{mcgrath2025time,
  author = {Sean McGrath and Takuya Kawahara and Joshua Petimar and Sheryl L. Rifas‐Shiman and Iván Díaz and Jason P. Block and Jessica G. Young},
  title = {{Time-smoothed inverse probability weighted estimation of effects of generalized time-varying treatment strategies on repeated outcomes truncated by death}},
  year = {2025},
  eprint = {2509.13971},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.13971},
  howpublished = {arXiv},
}
Figure