Crossover Multi-Arm Experiment — Multi-Treatment Within-Subject Design

A standard (two-arm) crossover trial has participants experience exactly two treatments (A then B, or B then A). A crossover multi-arm experiment extends this to three or more arms, requiring careful sequence design (e.g., Williams design) to ensure balance across periods and sequences. The analytic complexity and total study duration both increase with k, but so does the informational yield per participant.

A Williams design is a specific set of treatment sequences for k arms that guarantees: (1) each treatment appears exactly once in each period, and (2) each treatment precedes and follows every other treatment the same number of times. This balance eliminates first-order carryover effects from the treatment estimates, making it the preferred choice for multi-arm crossover trials when complete balance is achievable.

Yes. With k = 4 arms a complete Williams design requires four periods and 4 sequences (or 8 for full balance). The total study duration grows proportionally — k periods plus k-1 washout intervals. For k ≥ 5, incomplete crossover designs (where not every participant receives every arm) are sometimes used to limit participant burden, though they require more careful analysis to separate treatment and carryover effects.

Dropouts after at least one period provide partial information that can be recovered under a missing-at-random assumption using mixed-effects models (MMRM), which use all available measurements without imputation. Participants who drop out before completing any period contribute no data. Prevention — careful participant selection, manageable washout schedules, and close follow-up — is more effective than statistical rescue.

Because each participant contributes k observations, sample sizes are much smaller than for a parallel k-arm trial. Power calculations for crossover designs use within-subject variance (not total variance) and are conducted using the crossover-specific formulas in Jones and Kenward (2003) or dedicated software. Practical trials with k = 3 arms have been conducted with as few as 12–24 participants when within-subject variability is low.