Is crossover the same as a before-and-after study?

No. A planned crossover compares interventions across sequences; a simple before-and-after comparison may confound intervention with time.

Must there always be a washout?

No. Its need and feasibility depend on the intervention, but residual effects must still be considered.

Does randomizing sequence solve carryover?

No. It helps balance order-related influences but cannot guarantee that prior treatment stops affecting later measurements.

AI Glossary / Statistics and probability / Intermediate

Crossover study

Definition

A crossover study compares interventions by having the same participants receive different interventions during successive periods. In a randomized two-treatment, two-period design, participants are assigned to sequences such as AB or BA. Comparisons within participants can reduce the influence of stable differences between people, provided that time and residual intervention effects are handled appropriately.

What it means

The important feature is that participants cross from one intervention to another according to a planned sequence. This differs from a parallel-group study, where different participants receive different interventions throughout the comparison.

A crossover study can be efficient because a person's baseline characteristics are shared across their measurements. It also creates challenges: the person may learn, recover, become tired, or retain an effect from the first intervention. The second measurement therefore cannot automatically be treated as another independent observation under identical conditions.

In AI-related research, the approach might compare two interfaces or assistance modes, but persistent learning can make crossover unsuitable for some questions.

How it works

Define the interventions, outcome, period length, and sequence allocation in advance. In an AB/BA design, randomize participants to receive A then B or B then A. Where a temporary effect can dissipate, an appropriate washout interval may be incorporated between periods.

Distinguish the direct treatment effect from a period effect, which reflects when measurement occurs, and carryover, which reflects an earlier intervention's continuing influence. Counterbalancing order helps separate a common period effect from treatment, but does not automatically remove differential carryover.

Analyze repeated measurements with methods matching the design. Depending on the outcome and assumptions, this may involve a model with treatment, period, sequence, and participant effects. Plan missing-data handling before seeing results. A nonsignificant carryover test is not proof that meaningful carryover is absent.

Three practical examples

A usability team compares two temporary display layouts using equivalent tasks, randomizing order and recording completion time and errors for each participant.
Researchers compare reversible equipment settings on the same operators, allowing suitable recovery between sessions and accounting for measurement period.
An AI interface experiment compares assistance modes but first investigates whether experience with one mode permanently changes how participants use the other.

Think of it this way

Think of tasting two drinks in different orders, with a suitable pause between them. Each taster can compare both drinks, so personal preferences need not be inferred by comparing different people. But a lingering flavor from the first drink can alter the second rating. Changing the order is useful; it does not make a lingering flavor disappear.

Worked example

Consider an original illustrative study of two interface settings. Four participants are allocated equally to AB and BA sequences. Completion times in minutes are: participant 1, AB, A = 14 and B = 11; participant 2, AB, A = 18 and B = 15; participant 3, BA, B = 13 and A = 14; participant 4, BA, B = 17 and A = 18.

Within-person B-minus-A differences are -3,-3,-1,-1, averaging -2 minutes. Suppose these illustrative values were generated by B reducing time by two minutes and a common second-period practice improvement of one minute, with no carryover or treatment-by-period interaction.

In AB, B receives both benefits, giving -3. In BA, A receives the practice improvement, making B-minus-A equal -1. Equal representation of the two sequences averages away that common period contribution: (-3 - 1)/2 = -2.

The arithmetic shows why order matters. It is not a significance test or a realistic estimate of sampling variability. If only AB had been observed, the three-minute difference would mix the stipulated treatment and period effects.

Illustrative case study

In this fictional case, Orbit Workspace compares two AI-assisted document interfaces. Each tester uses both versions on comparable documents, and the team randomizes order. It initially expects crossover to provide a quick answer with fewer testers.

A pilot reveals that the first interface teaches a document-checking strategy that participants retain when switching. A short break does not erase that knowledge. The team therefore uses separate groups for its main question about first-time use, while retaining a crossover exercise for a different question about experienced users' preferences.

The invented example illustrates matching study design to the intended effect. It does not establish that either interface is superior or that crossover is generally unsuitable for software evaluation.

Common misconceptions and mistakes

Receiving both interventions does not guarantee an unbiased comparison. Differential carryover, changing conditions, dropout, and inconsistent tasks can distort results.

Do not call a break a successful washout without a substantive reason to believe the relevant effect has dissipated. Learned skills or irreversible changes may persist indefinitely for the study's purposes.

Nor does the repeated-measures structure justify analyzing all records as independent. The number of measurements is not the number of independent participants. Efficiency depends on within-person variation, the design, and the assumptions, rather than a universal rule that crossover always needs half as many people.

Business use

Businesses can use crossover studies when comparing temporary, reversible experiences and when individual differences would otherwise make comparisons noisy. Define whether the decision concerns first use, trained use, short-term performance, or sustained adoption.

Budget for the full participation period and assess the risk that people leave before completing all conditions. Record sequence, timestamps, task versions, and deviations so analysts can reconstruct what occurred. If an intervention changes future behavior permanently, a parallel design or another evaluation strategy may better match the business question.

Check your understanding

Check your understanding: Why is the AB difference -3 in the example when the stipulated direct effect is -2?

Answer: It includes a one-minute second-period improvement.

What makes averaging the two sequence contrasts recover -2 here?

Answer: Equal weighting and the stated common-period, no-carryover assumptions.

Would a permanent training benefit reliably disappear after a short rest?

Answer: No; it could make the crossover comparison unsuitable.

Related terms

Counterbalancing
Carryover effect
Repeated measures
washout period
period effect
randomization
parallel-group design
within-subject design

Learn next

Compare period effects with differential carryover using a two-sequence table. Then examine how a repeated-measures analysis accounts for participant variation and how missing second-period observations affect interpretation.

Sources

Sources: Lim and In, Considerations for crossover design in clinical study, https://doi.org/10.4097/kja.21165; the authors' article was read through its accessible archived PDF. Penn State STAT 509, Crossover Designs, https://online.stat.psu.edu/stat509/Lesson12, indexed carryover and washout discussion. The calculation and fictional software case are original; no clinical recommendations are implied.

Previous: Crossover point Next: Crout matrix decomposition