Multiple Imputation Methods for Handling Missing Data under Treatment Policy Strategy: A Scoping Review

In the world of medical research, the gold standard for proving a new drug works is the randomized controlled trial. In these studies, volunteers are assigned to receive either the new treatment or a standard one, and researchers track their health over time to see who improves. However, real life is messy. Participants often stop taking their medication, switch to a different treatment, or drop out of the study entirely due to side effects or personal reasons. These interruptions are known as intercurrent events. For decades, statisticians debated how to handle the data from these participants once they left the treatment path. Should the analysis pretend they never stopped, or should it only count the time they actually took the drug? A major update to international clinical trial guidelines introduced a specific approach called the "treatment policy strategy." This approach asks a simple, practical question: what happens to a patient if they are assigned to a treatment, regardless of whether they stick with it? It treats the assignment itself as the key factor, analyzing all outcomes as they occur, even after a patient has stopped the drug.

The challenge arises when a patient stops the treatment and then stops showing up for follow-up visits. Their health data becomes missing. If researchers simply ignore these missing numbers, the results of the trial can become skewed, potentially making a drug look better or worse than it truly is. To fix this, statisticians use a technique called multiple imputation. Imagine a puzzle where some pieces are lost; instead of leaving a gap, you create several different, plausible versions of the missing pieces based on the patterns you see in the pieces you still have. You then solve the puzzle with each version and combine the results to get a reliable answer. While this method works well when data is missing at random, it often fails under the treatment policy strategy because patients who stop treatment often change their health trajectory in ways that differ from those who stay on it. Standard assumptions no longer apply, and the missing pieces need to be filled in with more sophisticated logic.

A team of researchers at Imperial College London set out to map the landscape of these advanced solutions. They conducted a comprehensive review of the scientific literature, searching for every study that proposed or compared methods for filling in missing data specifically for the treatment policy strategy. They looked at papers published from the beginning of database records up to early 2026. After a rigorous screening process that removed duplicates and irrelevant studies, they identified 52 key articles that focused on this specific problem. Their goal was not to test a new drug, but to organize the mathematical tools researchers use to ensure clinical trial results remain honest and accurate when patients drop out.

The review revealed that the field has evolved significantly, particularly since the new guidelines were formally introduced in 2019. Before that, similar ideas existed but were described using different names, such as "de facto" or "effectiveness" estimands. The new guidelines helped standardize the language, leading to a surge in publications that explicitly addressed the treatment policy strategy. The researchers found that most of the work in this area comes from the United States and the United Kingdom, with the majority of papers being methodological studies designed to test and refine statistical models rather than reports of actual clinical trials.

The core of the review focused on four main families of methods used to handle these missing data points. The most common approach, found in about three-quarters of the studies, is called reference-based imputation. This method assumes that if a patient stops their assigned treatment, their future health outcomes will begin to look like those of a specific comparison group, often the patients who received a placebo or a standard care drug. There are several variations of this idea. One version, known as "jump to reference," assumes the patient immediately loses the benefit of the new drug and instantly adopts the trajectory of the comparison group. Another, called "copy increments in reference," assumes the patient keeps the benefit they gained up to the moment they stopped, but then their health changes at the same rate as the comparison group. A third variation, "last mean carried forward," assumes the patient maintains the average benefit they had at the moment they left the study.

The second major family of methods is called retrieved dropout imputation. This approach takes a different view. Instead of borrowing data from a comparison group, it relies entirely on the actual data collected from other patients who stopped the treatment but stayed in the study. If a patient drops out, their missing future data is estimated based on the observed patterns of those who also stopped treatment but continued to provide updates. This method is powerful because it uses real-world evidence from the trial itself, but it requires that enough patients who stopped treatment actually remain in the study to provide a reliable pattern. The researchers found that this method has been developed in various forms, some looking only at the last visit before a patient dropped out, and others using a more complex model that tracks the patient's entire history of treatment and stopping.

The review also identified two other, less common strategies. One is the "return-to-baseline" method, which assumes that after a patient stops treatment, their health reverts to the level it was at the very start of the study, with some random variation added. The other is the "washout" method, which is similar to the reference-based approach but is applied only to the final measurement of the study. It assumes that any benefit from the experimental drug disappears over time for those who stop, leaving only the effect of the placebo or standard care. In addition to these four main categories, the researchers found a growing number of "hybrid" methods. These are newer techniques that try to combine the strengths of the different approaches, such as using the real data from dropouts when available, but switching to a reference-based assumption if there isn't enough data to make a reliable estimate.

The researchers also examined how these methods are applied to different types of health outcomes. While the majority of the work focused on continuous measurements, such as blood pressure or pain scores, the review showed that statisticians have successfully adapted these techniques for binary outcomes (like whether a patient recovered or not), time-to-event data (like how long it took for a disease to return), and even complex counts of recurring events. However, the review highlighted a significant gap: while the methods for continuous data are well-developed, the tools for other outcome types are still maturing. Furthermore, the researchers noted that while many papers provided the computer code needed to run these imputation methods, fewer provided the code used to generate the simulated data for testing them. This makes it harder for other scientists to fully understand how the methods were validated.

A critical issue discussed in the review is the problem of "uncongeniality." This occurs when the mathematical model used to fill in the missing data does not perfectly match the model used to analyze the final results. This mismatch can lead to incorrect estimates of how certain researchers should be about their findings. The review found that while the standard statistical rules for combining results are widely used, some experts argue they may not be the best choice for these specific methods. Instead, some researchers suggest using alternative calculation methods, such as bootstrapping, to get a more accurate measure of uncertainty. The review did not declare one method superior to all others; rather, it emphasized that the choice of method depends heavily on the specific assumptions a researcher is willing to make about what happens to a patient after they stop treatment.

Ultimately, this scoping review serves as a vital map for the complex terrain of modern clinical trial analysis. It shows that while there is no single, perfect way to handle missing data when patients stop treatment, there is a robust and growing toolkit of methods available. The researchers concluded that understanding the specific assumptions behind each method is essential for choosing the right one. As clinical trials continue to evolve to reflect real-world conditions, the ability to accurately fill in the gaps left by missing data will remain a cornerstone of trustworthy medical science. The work of these statisticians ensures that when a doctor prescribes a treatment, the evidence supporting it accounts for the reality that patients may not always follow the plan perfectly.