Missing Data in a Trial: Why How It Is Handled Can Change the Result

Missing data are almost never truly random, so the method used to handle them can move a trial's result in either direction. The strongest analyses prevent missing data by design, use approaches like multiple imputation that carry the uncertainty forward instead of pretending a single guessed value is known, and test whether conclusions hold under less favorable assumptions. When you read a trial, look for how much data were missing, why, and whether the authors ran a sensitivity analysis.

Three ways data go missing

Not all missing data are the same, and the label a statistician assigns changes which analysis is honest. Data missing completely at random means the reason a value is absent has nothing to do with anything, observed or unobserved; the gaps are just a random thinning of the sample. This is the friendly case, and it is rare.

Missing at random is a subtler idea and a poor name. It means that once you account for the information you did observe, the chance a value is missing no longer depends on the value that would have been recorded. Missing not at random is the hard case: a measurement is absent for reasons tied to what it would have shown, such as patients leaving precisely because they were getting worse. Real trials usually sit somewhere between the last two, which is why the handling method is not a technicality.

Why complete case analysis and LOCF mislead

The two most tempting shortcuts are also the most misleading. Complete case analysis simply drops anyone with a missing value. That throws away information, shrinks the sample, and, when data are not missing completely at random, bends the result because the people who stayed differ systematically from the people who left.

Last observation carried forward takes a patient's final recorded value and pretends it held for the rest of the trial. It looks tidy, but it invents certainty that does not exist and can push a result in either direction depending on why people dropped out. Its deeper flaw is that it treats a guessed value as if it were measured, which shrinks the standard error and makes the finding look more precise than the data support. The panel convened by the National Research Council discouraged leaning on either method as a primary analysis.

What multiple imputation actually does

Multiple imputation takes the uncertainty seriously instead of hiding it. Rather than filling each gap with one best guess, it builds several complete datasets, each with slightly different plausible values drawn from a model of the observed data. Each dataset is analyzed on its own, and the results are pooled using a set of rules that combine the estimates and, crucially, widen the uncertainty to reflect how much the datasets disagree.

The payoff is a treatment estimate whose confidence interval honestly includes the fact that some values were never seen. Multiple imputation is valid when the missing at random assumption is reasonable, which is why it is usually paired with a check of what happens if that assumption fails.

The estimand question: what are we estimating?

Before arguing about imputation, careful trials now ask a more basic question: what exactly is the target? The estimand framework, set out in the addendum known as ICH E9(R1), pushes researchers to define the treatment effect precisely, including how they will handle events that occur after randomization and disturb the picture, such as a patient stopping the drug or taking a rescue medication.

This matters because the same trial can answer several different questions. Do we want the effect of being assigned the treatment, whatever people did next, or the effect of actually staying on it? Naming the estimand first turns missing data from an afterthought into a design decision, and it clarifies which absent values are a nuisance to impute and which reflect a real part of the question being asked.

Sensitivity analysis: does the answer survive a worse assumption?

Because you can never prove from the data alone that values are missing at random, the honest move is to test whether the conclusion depends on that hope. A sensitivity analysis re-runs the result under deliberately less favorable assumptions, for example that patients who dropped out fared worse than their observed data would suggest.

If the treatment effect holds up under these gloomier scenarios, the finding is robust. If a modest, plausible shift in assumptions erases it, the result was resting on the missing data behaving nicely, and it deserves caution. A trial that reports a sensitivity analysis is showing its work; one that does not, next to a lot of missing data, is asking for trust it has not earned.

What to look for as a reader

You do not need to run the models to judge them. Find how much data were missing and in which arm, because a large or lopsided amount of dropout is a warning regardless of the method. Check whether the trial prespecified its handling of missing data or appears to have chosen the approach after the fact.

Then look for two words: sensitivity analysis. Their presence tells you the authors asked whether their answer survives a pessimistic assumption. Their absence, sitting beside a lot of missing data, is often the most informative thing on the page.