Menu

New methodology guidance on inverse probability of censoring weighting in clinical trials

18 Nov 2025

A new study from our methodology researchers and collaborators provides practical guidance on how to improve the use of the inverse probability of censoring weighting (IPCW) method in clinical trials. The paper, recently published in Statistical Methods in Medical Research, explains how to apply IPCW effectively and how to stabilise IPCW weights to improve precision, with awareness of possible added bias.

Clinical trials often encounter post-randomisation events such as patients stopping treatment, switching, or use of rescue therapy. These events, which complicate the analysis and interpretation of trials results, are known as intercurrent events (or protocol deviations in a broader sense).

To address these complexities, estimands provide a structured way to define the precise question a trial aims to answer. A hypothetical estimand, for example, addresses intercurrent events by asking “What would the treatment effect look like if certain intercurrent events had not occurred?”

IPCW is a statistical method that targets hypothetical estimands. It works by:

  • Giving more weights to participants who do not experience intercurrent events, so that they represent similar people who experience the event and therefore are censored in the analysis.
  • Deriving these weights from each participant’s probability of experiencing the intercurrent event, the less likely the event, the larger the weight.

These weights can become very large or highly variable, which can lead to unstable estimates. Using weight stabilisation, we can reduce this variability and improve precision.

Our work focuses on how to use IPCW effectively when estimating hypothetical estimands in clinical trials. Although IPCW is a well-established method, its practical application has been limited by concerns about its precision and robustness. Our previous work demonstrated that IPCW is a strong and reliable alternative to per-protocol analysis.

Building on that, this new study uses an illustrative example and systematic simulation studies to see how IPCW works with unstabilised weights and with different stabilisation methods. These approaches are evaluated under both correctly specified and mis-specified substantive outcome models. The findings give practical advice on how to stabilise weights appropriately and clarifies the bias-variance trade-offs involved.

Key findings

  • Compared to unstabilised IPCW, stabilised weights often improved efficiency (smaller standard errors, narrower confidence intervals), especially when the stabilisation incorporated baseline covariates.
  • When the substantive analysis model was mis-specified, stabilised weights risk increasing bias. Improved precision did not always come without cost.

What this means for applied researchers

For analysts who are designing trials or planning analyses that target hypothetical estimands, the key message is: consider stabilising IPCW weights but do so thoughtfully. 

The detailed recommendations include:

  • The numerator of the weights should include functions of baseline covariates and/or time when there is reason to believe those covariates or time influence censoring by the intercurrent event.
  • When there is low risk of mis-specification for substantive outcome model, stabilisation poses little risk of introducing additional bias and is often beneficial for improving precision.
  • When there is substantial risk of mis‐specification (e.g., complex time‐varying effects) for substantive outcome model, using unstabilised weight can be a safer choice to avoid added bias. Conducting sensitivity analyses with stabilised weights can provide useful context, and each result should be clearly reported.

Further information: