How to Read a Crossover Trial: Washout, Carryover, and Paired Analysis

A crossover trial gives each participant every treatment in a randomly assigned order, so each person acts as their own comparison and the study can detect differences with fewer people. That efficiency comes with a specific risk: the first treatment can leave effects that spill into the next period, which is why you look for an adequate washout and a paired analysis. Read one by checking that the design suits a stable, reversible condition and that the analysis respects the within-person structure instead of treating the periods as separate groups.

What a crossover trial is trying to do

A parallel trial splits people into groups and compares one group against another. A crossover trial takes a different route: every participant receives each treatment, one after another, and the order is assigned at random. Because the same person supplies the data for both treatments, the comparison is made within that person rather than between strangers who happen to differ in age, genetics, or disease severity.

That within-person comparison is the whole point. Much of the noise in a trial comes from how different people are from one another. Remove that source of variation and you can detect a genuine treatment difference with far fewer participants. For a stable condition and a treatment that works quickly and wears off, this makes the crossover design remarkably efficient.

Carryover, the flaw the design is built around

The efficiency has a catch. If the first treatment is still acting when the second period begins, the second period is not measuring the second treatment cleanly. This spillover is called carryover, and it is the central worry in any crossover trial.

The defense is a washout period, a gap between treatments long enough for the first one to clear and its effect to fade before the next begins. When you read a crossover trial, find the washout and ask whether it is plausibly long enough given how the treatment works. A drug with a long half-life or a lasting biological effect needs a longer washout. The reporting guidance is blunt about a related trap: in the common two-period design, carryover cannot be reliably detected after the fact, so the trial has to prevent it by design rather than test for it later.

Why the analysis has to be paired

Since each participant contributes a result on both treatments, the two numbers are linked. The right analysis compares each person's result on one treatment with their own result on the other, then summarizes those within-person differences. This is what people mean by a paired analysis.

Getting this wrong is a real and detectable error. If a trial analyzes the periods as if they were two separate groups of people, it throws away the pairing that made the design efficient in the first place and reports more uncertainty than the data actually contain. When you read the methods, check that the analysis accounts for repeated measurements in the same individual and reports the treatment difference with a confidence interval built on paired data.

When a crossover design fits, and when it does not

Crossover trials work well for chronic, stable conditions where the treatment acts fast and is fully reversible: think of a bronchodilator for stable asthma, or a drug for a steady symptom that returns once the drug stops. The condition needs to sit still long enough that the two periods are comparable.

The design fails when those assumptions break. If the illness improves or worsens on its own during the trial, the periods are no longer a fair comparison. If a treatment cures the condition, changes it permanently, or takes weeks to act and weeks to leave, there is no clean second period to measure. A treatment that could be dangerous to stop and restart also rules out the design. Seeing a crossover trial used for an unstable or irreversible condition is itself a warning sign.

A short checklist for reading one

Start with the title and abstract: a well-reported crossover trial says it is one, because the design changes how you should interpret everything that follows. Then look for four things. Is the condition stable and the treatment reversible, so the design is appropriate at all? Is there a washout, and is it long enough for this treatment? Does the sample size calculation account for within-person variability rather than borrowing a parallel-trial formula? And is the analysis paired, with participant flow and baseline characteristics shown by sequence and period?

If those pieces are present and sensible, the trial is using the design honestly. If the washout is missing, the analysis is unpaired, or the condition is one that drifts over time, treat the efficiency of the design as a liability rather than a strength.