Why Treatment Order Matters
In a crossover study, each participant receives more than one treatment, usually in separate periods. A participant might receive treatment A first and treatment B second, while another receives B first and A second. Because participants serve as their own comparison, stable differences between people are controlled by the design. But the comparison can be misleading if outcomes change over time or if an earlier treatment continues to affect a later period.
The central question is not simply whether each participant received both treatments. It is whether the observed difference can reasonably be attributed to the treatment itself rather than to when it was given or to the treatment that came before it.
For two treatments, A and B, the two common sequences are AB and BA. Randomly assigning participants to these sequences helps prevent systematic differences between the order groups. It does not guarantee that treatment effects have ended before the next period begins. A washout interval may be used between periods, but its adequacy depends on the treatment and outcome; merely scheduling a gap does not prove that carryover is absent.
A Model for the Two-Period Comparison
A simple additive model helps separate the sources of change. Let \(Y_{ij}\) be participant \(i\)'s outcome in period \(j\), and write
Here \(b_i\) represents the participant's stable baseline level, \(\pi_j\) is the effect associated with period \(j\), and \(\tau_{T_{ij}}\) is the direct effect of the treatment given in that period. The carryover term \(\kappa\) is zero in the first period and may be nonzero later. The remaining term \(\varepsilon_{ij}\) represents other variation not captured by these components.
This model is a way to reason about the design, not a guarantee that real outcomes follow an additive pattern. A simple within-participant treatment comparison is easiest to interpret when the direct treatment contrast is reasonably stable across periods, period changes are handled by the design, and there is no meaningful carryover. If treatment effects depend on period or sequence, a single overall contrast may conceal important differences.
How Period Effects Can Cancel—and Carryover May Not
To see the role of order, define the treatment contrast as B minus A. In an AB participant, this is the second-period outcome under B minus the first-period outcome under A. In a BA participant, it is the first-period outcome under B minus the second-period outcome under A. Suppose the direct B-minus-A effect is \(\tau_B-\tau_A\), and the period difference is \(\pi_2-\pi_1\). Then the expected contrasts under the two sequences are
The signs differ because period 2 is paired with B in sequence AB but with A in sequence BA. If the two sequences are equally represented, averaging their expected contrasts gives
Thus, when there is no carryover, equal representation of the two sequences cancels an additive period effect in this simple model. But if carryover differs by the preceding treatment, the averaged contrast can still be biased. With unequal sequence sizes, a simple pooled average of participant contrasts does not necessarily cancel the period effect exactly. Random assignment can help balance groups on average, but it does not remove a systematic carryover effect.
Worked Example: Reading a Balanced Two-Sequence Study
Suppose a hypothetical study compares A with B on a symptom score, where lower scores are better. Three participants are assigned to each sequence. The table shows each participant's score under A and B, along with the period in which each treatment was received.
| Sequence | Participant | Period 1 | Period 2 | B minus A |
|---|---|---|---|---|
| AB | 1 | A: 12 | B: 9 | 9 − 12 = −3 |
| AB | 2 | A: 10 | B: 8 | 8 − 10 = −2 |
| AB | 3 | A: 11 | B: 8 | 8 − 11 = −3 |
| BA | 4 | B: 9 | A: 11 | 9 − 11 = −2 |
| BA | 5 | B: 8 | A: 10 | 8 − 10 = −2 |
| BA | 6 | B: 10 | A: 11 | 10 − 11 = −1 |
For AB, the mean B-minus-A contrast is
For BA, the mean contrast is
With equal numbers in the two sequences, the overall mean contrast is
The observed B-minus-A difference is about \(-2.17\) symptom-score points, so scores were lower under B on average in these hypothetical data. The sequence-specific means differ by 1 point. That difference is a reason to examine whether time, order, or carryover could matter; it is not, by itself, proof of a particular effect. The design and subject-matter knowledge are needed to judge whether the overall contrast has a defensible treatment interpretation.
Period Effects: Change That Is Not the Treatment
A period effect can arise from natural recovery, disease progression, learning, fatigue, seasonal change, or another influence that differs between study periods. If every participant receives A first and B second, treatment and period are completely entangled: any observed difference could reflect the treatment, the later period, or both. Having each participant receive both treatments does not resolve this problem when the order is fixed.
Worked Example: Fixed Order Confounds Treatment and Time
In a hypothetical symptom study, four participants receive A in the first period and B in the second. Their scores are 20 then 16, 18 then 14, 22 then 18, and 16 then 12. The B-minus-A differences are
The scores fell by 4 points between periods for every participant. But because A was always first and B always second, the data alone cannot show that B caused the decrease. For example, if symptoms would have fallen by 4 points over time even without switching treatment, the observed difference is entirely compatible with no direct treatment effect. Randomizing participants to both AB and BA sequences would help separate a common period change from a treatment contrast.
Carryover: The First Treatment Reaches into the Next Period
Carryover is different from a general period effect. It is specifically the lingering influence of an earlier treatment. It is a concern when a treatment's biological action lasts beyond its scheduled period, when its effect builds slowly, or when it changes the participant in a way that persists after the treatment stops. A washout period is useful only if it allows the relevant effect to dissipate sufficiently for the next period's outcome to represent the next treatment.
Worked Example: Residual Benefit Distorts the Contrast
Suppose a hypothetical study compares A with B using sleep latency in minutes, with lower values indicating falling asleep sooner. Assume the true direct effect of B is \(-2\) minutes relative to A, and that the second period raises latency by 1 minute for everyone. Also suppose A has a residual carryover effect that lowers the next period's latency by 3 minutes, while B has no carryover effect. Start from a baseline latency of 20 minutes.
For a participant in sequence AB, period 1 under A is 20 minutes. In period 2, the direct effect of B, period change, and A carryover combine:
The observed B-minus-A contrast in this sequence is \(16-20=-4\) minutes. For a participant in sequence BA, period 1 under B is \(20-2=18\) minutes. Period 2 under A is \(20+1=21\) minutes, since the assumed B carryover is zero. The contrast is \(18-21=-3\) minutes. Averaging the sequence contrasts gives
The direct B-minus-A effect in the stated model is \(-2\) minutes, but the balanced observed average is \(-3.5\) minutes because A's carryover makes the following B period look especially favorable. The period effect cancels in the balanced average, but this treatment-specific carryover does not.
When a Crossover Design Is Suitable
Crossover designs can be efficient when outcomes are reasonably stable over the study, treatments act for a limited time, and participants can safely receive each treatment. The within-person comparison controls stable participant characteristics, but it does not automatically control changes over time or lingering effects. Before choosing the design, investigators should ask whether the clinical question and treatment allow those assumptions to be credible.
If recovery, worsening, or another time trend is expected, period effects may complicate interpretation.
If meaningful effects remain after stopping treatment, later-period outcomes may reflect both current and previous treatments.
Randomizing participants to sequences helps guard against systematic order-group differences; it does not eliminate carryover.
Each participant must be able to receive each treatment, and delaying or stopping an effective treatment must be acceptable.
If these conditions are doubtful, a parallel-group design, in which different participants receive different treatments, may be more appropriate. The choice depends on the scientific question and the treatment's time course, not simply on a desire to use each participant as their own control.
Common Mistake
A common mistake is to reason, “Everyone received both treatments, so each participant is their own control and order cannot matter.” The within-participant comparison does control stable individual differences, but it does not make the periods interchangeable. When treatment order is fixed, a time trend is confounded with treatment; when effects persist, the second treatment may be contaminated by the first. Randomized sequences help address order and additive period effects, while adequate washout and subject-matter reasoning are needed to assess carryover.
Check Your Understanding
Use the design logic in this tutorial to answer the following questions.
- In a two-treatment crossover study, what distinguishes a period effect from a carryover effect?
- Why does assigning every participant to A first and B second make treatment and period difficult to distinguish?
- Under the additive model, why can equal representation of AB and BA cancel a common period effect in the average contrast?
- A treatment's effects last into the next period despite a short gap. What problem does this create for interpreting the later outcome?
- Why does random assignment to treatment sequences help, and what important problem does it not necessarily solve?