A New Way to Impute Missing Data in Clinical Trials Targeting Treatment Policy Estimands

When patients stop taking their assigned treatment in a clinical trial - whether due to side effects, lack of effectiveness, or personal choice - researchers often still want to understand how that treatment performs in the real world. That’s the goal of the treatment policy estimand strategy in the ICH E9(R1) Addendum: to assess the treatment effect regardless of treatment adherence.

But there’s a big problem: patients who withdraw from treatment often stop attending follow-up visits too, leading to missing outcome data. This makes it hard to estimate the treatment policy effect. So how can we fill in the blanks in a statistically valid way?

There are three main ways statisticians can currently impute missing data post-treatment withdrawal in longitudinal trials in this context:

1. Ignore treatment withdrawal status, assume all missing data, including post-treatment withdrawal data are Missing At Random (MAR), and proceed using standard Multiple Imputation. This is a simple approach, but often the assumptions are unrealistic and don’t fully align with the treatment policy strategy.

2. Use Reference-Based Multiple Imputation (RBI), which assumes that patients who stop treatment behave like those in a specified reference group (typically control/placebo). Patient data in the active treatment arm is imputed based on a model fitted to the specified reference arm data.  This approach uses no actual post-treatment withdrawal data for deviators during imputation, even when some is available. But this means it can be implemented with none or limited observed post-treatment withdrawal data.

3. Use Retrieved Dropout (RD) Multiple Imputation, which builds a statistical model for imputation using both on- and off-treatment data to impute the missing values. The aim of using such an approach is that data missing post treatment withdrawal will be imputed based on the distribution of the observed post treatment withdrawal data. The problem? With limited (or no) post-treatment withdrawal data, such models often become unstable or even inestimable.

You can learn more about these three methods and how to implement in practise in session 5 (and practical 3) available here: OSF | MRC-NIHR-TMRP Workshop: Methods of Analysis for Different Estimands (MADE) 

Introducing a new method

Our recent paper proposes a new solution for this problem - Retrieved Dropout Reference-Based Centred Multiple Imputation (RD RBI-C). This novel method blends RBI and RD into a single, flexible imputation approach that is able to realise the benefits of both separate methods.

Here's the underlying  idea:

1.    Start with a core reference-based imputation (RBI) model, which borrows the

distribution of outcomes from a specified reference group (e.g. control/placebo).

2. Then extend it with a series of additional parameters so that the extended model becomes equivalent to a retrieved dropout (RD) type of model that uses actual post-treatment withdrawal data -if available- to make more informed imputations. 

3. The extended model is fitted using a Bayesian framework with typically uninformative priors for the core reference-based model parameters and mildly informative priors for the additional parameters. This prevents the extended model from falling apart when observed post-treatment withdrawal data is scarce, while still leveraging the data that is available to inform imputations.

   
   

This new method will result in imputed data that behaves similar to imputed reference-based MI data when there is none/limited observed post-treatment withdrawal data. But when there is a large proportion of observed off‐treatment data available this will over power the mild priors for the additional parameters and the imputation will appealing behave like the chosen retrieved drop out, off‐treatment,  model.

Benefits of this new approach  

Our new approach avoids strong model assumptions based on data in opposing treatment arms by utilising observed off-treatment data when it is available.

It preserves alignment with the treatment policy strategy.

As it is implementable regardless of the proportion of realised missing data, this approach enables a non-adaptive, pre-specified analysis plan that is more transparent than adjusting your analytical approach based on the data you end up with.

Overall, it offers a statistically principled way to address a common real-world missing data challenge in clinical trials targeting treatment policy estimands.

Want the full details?

If you're working on clinical trials or involved in regulatory statistics, we encourage you to dive into the full research article. You'll find the detailed methodological algorithm with examples, providing an overview of how to implement this approach -even when a large proportion of your data is missing. We provide open-access SAS code to facilitate implementation.

Read the full paper here: Handling Partially Observed Trial Data After Treatment Withdrawal: Introducing Retrieved Dropout Reference‐Base Centred Multiple Imputation

And if you are heading to PSI 2025 you can hear all about this new methods there - check out the full conference Programme.