A randomized crossover trial assigns participants to an order of interventions. In a simple two-treatment design, one sequence receives A then B and the other B then A. Participants with observations under both treatments can contribute a within-person comparison. That can improve precision, but only when the condition, treatment effects, follow-up and analysis fit the design.[4][5]
A later treatment change does not by itself establish a randomized crossover. Rescue treatment, nonadherence, switching during ordinary care and an open-label extension are different from randomized assignment to a planned sequence. If a report does not describe sequence allocation clearly, that is an information gap, not proof that the order was randomized or that it was nonrandomized.
Five units to keep separate
| Unit | Meaning in a simple AB/BA trial | What to check |
|---|---|---|
| Participant | A person followed across periods | Numbers randomized, observed and analyzed |
| Sequence | Assigned order, A then B or B then A | How the sequence was generated and concealed |
| Period | Interval for a treatment and its outcome assessment | Timing, duration and comparable measurement |
| Treatment | Intervention assigned in that period | Actual exposure and deviations from assignment |
| Observation | A measurement linked to person, treatment and period | Correlation and missing measurements in the analysis |
Twenty people observed twice are still twenty participants, not forty independently randomized people. For a cluster crossover, the randomized unit is a group, and analysis must address both repeated periods within groups and similarities between people in the same group. This guide mainly concerns individually randomized crossover trials.[4]
Why pairing can improve precision
A person’s relatively stable characteristics may affect their response under both treatments. Comparing A with B within that person can remove some between-person variation. For a simple continuous outcome, the important variability is the spread of the individual A-minus-B differences.[4][5]
When the two responses are positively correlated and complete paired data support the comparison, ignoring the pairing generally makes the standard error larger and the confidence interval too wide. It loses a potential precision advantage; it does not automatically make the result look more certain. Cochrane identifies this particular unpaired analysis as a unit-of-analysis error that can underweight a crossover trial in a meta-analysis.[4]
This direction is not universal. Negative correlation, incomplete observations, other outcomes, additional periods and other model errors require their own assessment. The familiar problem of ignoring positive correlation within randomized clusters is not interchangeable with ignoring pairing in a within-person treatment contrast. Neither repeated observations nor a narrow interval establishes an unbiased result.
A useful question is therefore: “Was uncertainty calculated for the treatment contrast and dependence structure actually studied?” See the guide to confidence intervals and p values for why an estimate and its interval matter more than a significance label alone.
Period effects are not carryover
A period effect is a change over time that is not the treatment contrast itself: symptoms may evolve, seasons may change, background care may shift, or participants may become familiar with a test. In a balanced AB/BA trial, period effects can cancel in a suitable analysis if the sequence balance is retained and missingness is not differential. Initial equal allocation is not enough when different people are lost from the two sequences. Period adjustment may be needed.[4]
Carryover is an earlier treatment’s residual influence on a later period. The intervention might remain in the body, or its biological or behavioral effects might persist after the intervention itself has disappeared. An observed order-dependent pattern need not identify which mechanism caused it.
ICH E9 makes a precise warning about the two-treatment, two-period design: carryover cannot be statistically distinguished from a treatment-by-period interaction, and the corresponding test has low power. A nonsignificant test therefore does not prove that carryover is absent. More elaborate designs do not make the concern disappear automatically.[5]
A washout is a rationale, not a certificate
A washout is intended to allow earlier effects to resolve. Its adequacy depends on the intervention, outcome and timing of the next assessment, not simply on naming a gap between periods. Drug elimination and disappearance of the relevant effect can occur on different timescales. Withdrawal or rebound during a gap can create another difficulty.[4][5]
Prior evidence should support a sufficiently stable condition, an effect that develops within each treatment period, and reversibility before the later outcome is measured. A distinct long gap is not always essential if enough time passes before that outcome; conversely, a stated washout is not sufficient evidence that the required separation occurred. These are design-appraisal principles, not instructions for choosing a treatment schedule.
| Reported pattern | Reasonable interpretation | Unsupported leap |
|---|---|---|
| A washout was included | A measure intended to reduce residual effects | Carryover was eliminated |
| No statistically significant carryover test | The test did not detect the specified pattern | No meaningful residual effect exists |
| Similar estimates in both periods | Some consistency in the observed data | Proof of no carryover or interaction |
| More withdrawals after one treatment | Retention may depend on exposure or response | Complete pairs remain a random subset |
First-period-only analysis needs its own justification
Choosing to discard later periods only after a preliminary carryover test is a criticized two-stage strategy. The test can be influenced by initial differences between sequence groups, and using it to choose the final analysis can introduce bias. The first period is not an automatic repair selected by a p value.[4]
A different decision is to plan a first-period-only analysis in advance for a sound scientific reason, for example, a review’s judgment that the crossover design is unsuitable for the outcome. That comparison is effectively parallel-group and gives up the intended within-person information. It still requires a clear rationale and suitable bias assessment.
Reviewers must also ask which first-period results are available. If trialists selectively report them after detecting carryover, a review restricted to available first-period data may inherit that selection. A prospective review rule does not remove selective availability in the underlying literature.[4]
Follow people through every period
A single final sample size can hide the most important losses. Trace each randomized sequence from allocation through treatment, assessment, any gap and the next period. Separate the number receiving treatment from the number contributing an outcome and the number included in each analysis. Record when withdrawals occurred and why.[4][5]
Participants who improve, worsen or experience adverse events may be less likely to contribute later observations. A simple complete-pair analysis then excludes people missing either measurement and may analyze a selected subset. Equal dropout percentages are not proof that the remaining comparison is unbiased. Nor does a “completer” label establish an intention-to-treat analysis.
No missing-data method is universally valid. The assumptions, patterns of missingness and sensitivity of the conclusion need explanation. Carrying an observation forward can be particularly problematic when it was obtained before an earlier treatment’s effect had resolved.[4][5] The missing-data and estimands guide and analysis-set guide explain related distinctions.
Harms need timing and denominators too. Summaries by treatment and period, together with the preceding exposure, are more informative than an unexplained pooled count. A later-period event may be difficult to attribute because an earlier treatment could still matter.[5]
Match the analysis to the design
A paired t test can be appropriate for a simple continuous-outcome crossover when carryover and period effects are not material concerns. It is not a universal method for every crossover. Other outcomes or designs may need models that address period, sequence and repeated observations differently.[4]
The report should identify the treatment contrast, outcome timing, analysis population, included model effects, handling of missing data and uncertainty estimate. Ask whether the chosen method matches the planned question and whether primary, supporting and exploratory analyses are distinguished.[5]
ICH E9 places the principal analysis features in the protocol. A more detailed statistical analysis plan can follow and can be refined during blinded review, with its finalization and unblinding dates documented. This is not a rule that every final analysis detail must predate the first participant. Changes to principal protocol features need appropriate documentation; data-dependent choices must not be quietly presented as original plans.[5] The registration and outcome-switching guide explains why dates and document versions matter.
For meta-analysis, a paired effect estimate needs a compatible uncertainty estimate. If a missing within-person correlation or difference-score variability is borrowed or assumed, explain that choice and examine plausible alternatives. Reporting an assumed correlation as an observed trial statistic would be misleading.[4]
Suitability comes before a favorable result
A simple crossover is often poorly suited to irreversible outcomes, curative or enduring effects, rapidly changing conditions, or follow-up in which many participants cannot contribute later observations. Randomization cannot reverse an earlier outcome or make a lasting treatment effect disappear.[4][5]
A positive short-term result should remain tied to the studied population, intervention, outcome and timing. It does not establish long-term benefit or safety, and it does not authenticate a separately sourced material. To apply these questions to an actual report, start with how to read a research study or browse the Study Library.
Reporting guidance and this review’s limits
The 2019 CONSORT crossover extension is a design-specific reporting document. The current official indexes identify general CONSORT 2025 and list the crossover extension under the CONSORT 2010 extensions. That listing does not establish a newly revised 2025 crossover checklist or prove that an individual trial is valid.[1–3][6]
This September 3, 2026 editorial review checked the official indexed PubMed abstract and metadata, Cochrane’s crossover sections, selected FDA E9 sections, the two official CONSORT indexes, and the material E9 pages and pairing equations. The BMJ route had returned an access restriction during supporting review. Its full article, full checklist and flowchart were not freshly reviewed here; the PMC link below is an identified access route, not a claim of full-text inspection. No exhaustive correction or retraction search, independent specialist review, or professional certification is claimed.
Cochrane’s page identifies Handbook version 6.5 (2024), with Chapter 23 last updated in October 2019. FDA E9 is the final September 1998 guidance. The date of this editorial check does not change those source versions.[4][5]
Frequently asked questions
Is every treatment switch a randomized crossover?
No. The design assigns participants to intervention sequences. Rescue use, nonadherence and a later open-label phase do not by themselves establish that design.
What is the difference between sequence and period?
Sequence is the assigned order of interventions. Period is the interval in which a particular intervention and its outcome assessment occur.
Does ignoring pairing always make uncertainty too small?
No. In the usual complete paired continuous comparison with positive correlation, it generally makes the interval too wide. Other data structures and model errors cannot all be assigned that direction.
Does a washout prove there was no carryover?
No. Prior knowledge and the timing of the later outcome must support reversibility. A named gap or nonsignificant carryover test is not proof.
Can investigators test carryover and then keep only period one?
That data-dependent two-stage strategy can bias the analysis. A scientifically justified plan to use first-period data from the outset is a different decision, with its own loss of information and reporting limitations.
Why does dropout matter so much?
Losing a later measurement removes part of the planned within-person comparison. If the loss relates to response, harm or sequence, complete pairs can be a selected subset.
Does complete CONSORT reporting mean the trial is trustworthy?
It helps readers see what was done. It does not certify suitable design, correct analysis, low bias, clinical applicability or the identity of another material.
Primary and official sources
- Dwan and colleagues, CONSORT extension to randomized crossover trials, BMJ 2019;366:l4378, original article route; access limitation described above.
- PubMed Central identifier and access route, PMCID PMC6667942, full text not freshly inspected.
- PubMed record, PMID 31366597, official indexed abstract and bibliographic identity.
- Cochrane Handbook Chapter 23, crossover sections 23.2.1–23.2.9; version and chapter date distinguished.
- FDA final ICH E9 landing page and its official guidance PDF, sections 3.1.2, 5.1, 5.3 and 5.5, with introduction and scope.
- Official CONSORT published-statements index and extensions index, current general statement and design-specific extension identity.
Last editorial evidence review: September 3, 2026. This is a source-based methods explanation, not individual medical or statistical advice. Send a correction or source update.