Extract Crossover Results Without Losing the Pairing

Confirm the design and question

In a crossover trial, participants receive more than one intervention in a planned sequence. The same person can contribute measurements under both conditions, so a table showing two treatment means does not establish independent groups. Extract sequence allocation, period length, washout, outcome timing and whether the condition is sufficiently stable for the design’s intended comparison.

Cochrane’s trial-variant guidance discusses crossover suitability, carry-over and appropriate analysis. Use it to identify what information is missing, rather than replacing a trial’s design with a familiar parallel-group formula.

Identify the reported analysis

Look for a within-person treatment contrast, its standard error or interval, and the number of participants contributing paired data. Record adjustment for period and sequence where reported. Distinguish participants randomized, participants completing each period and participants included in the treatment comparison.

A treatment-period mean and SD are insufficient to recover the variance of a within-person difference unless additional dependence information is available. Do not set the correlation to zero merely because it was not reported. If an approximation is necessary, state the assumed correlation and examine sensitivity to defensible alternatives.

Follow a hypothetical extraction

Suppose 30 participants are randomized, 28 complete period one and 26 provide outcomes under both treatments. The reported paired treatment difference is −2 units with a standard error of 0.8 units. Record 26 as the paired-analysis denominator, while preserving the randomized count of 30.

The inverse-variance weight associated with that reported standard error is 1/(0.8²) = 1.5625 on the relevant scale. It is not based on pretending that the two conditions contain 26 independent people each. An analysis using only period one answers a different data-selection question and may lose precision; it must be justified and documented rather than chosen after seeing the results.

Resolve design limitations before pooling

Check whether the intervention can have persistent effects, whether participants’ underlying condition changes across periods, and whether missing second-period outcomes depend on first-period experience. A washout label does not by itself establish absence of carry-over.

Keep the design-specific limitations alongside the effect estimate. If the review combines crossover and parallel trials, verify compatible outcome definitions and treatment contrasts, and describe how each variance was obtained. A usable extraction ends with a clear chain from the randomized sequences to the analyzed within-person comparison, including unresolved assumptions and exclusions.

Source verification: 6 November 2026. Examples are hypothetical.