Can several N-of-1 trials be combined?

Yes, with an appropriate analysis.

AI Glossary / Statistics and probability / Advanced

N-of-1 trial

Definition

An N-of-1 trial compares interventions within one individual across repeated periods. A randomized version assigns the intervention order by chance, allowing the individual to serve as their own comparison.

What it means

The primary question concerns that person's response under the studied conditions. Repeated measurements can provide useful evidence, but they do not create a large sample of independent people. An observational record of one person is not automatically a randomized N-of-1 trial.

How it works

Specify the interventions, outcome, period length, assignment scheme, and analysis before collecting results. Repeated crossovers can separate intervention comparisons from some sources of variation. Consider delayed effects, carryover between periods, changes over time, and dependence between nearby measurements.
This design is better suited to effects that can reverse within a workable period than to permanent changes. Clinical applications require appropriate clinical oversight; the nonclinical example below illustrates statistical reasoning rather than treatment decisions.

Three practical examples

A fictional user-experience study compares two reversible display layouts for one participant across repeated task periods.
A fictional accessibility evaluation compares two input settings for one user, recording both task performance and burden.
A fictional research training exercise compares blocked random assignment with a simple before-and-after sequence to expose time-trend bias.

Think of it this way

Instead of asking which option works better on average for a roomful of people, repeatedly compare options for the same person. The comparison still needs a fair schedule: later attempts might improve through practice even if the option itself makes no difference.

Worked example

Consider an original nonclinical trial of display layouts A and B for one participant. The outcome is the period's average task-completion time, with lower values preferred. Each period uses fresh tasks prepared to be comparable.
There are four blocks, each containing two consecutive periods. Within every block, independently flip a fair coin to decide whether the order is AB or BA. Do not impose an additional rule requiring exactly two blocks of each order.
Suppose the realized orders are AB, BA, BA, and AB. This happens to be balanced, but balance was not forced. The distinction matters for the randomization calculation below.
The four A-period averages are 12, 15, 11, and 14 time units. The corresponding B-period averages in the same blocks are 10, 12, 10, and 12. Subtracting B from A gives block differences 2, 3, 1, and 2.
Their mean is two time units in favor of B. The overall A average is 13 and the B average is 11. All four observed block differences favor B, but the number of independently randomized blocks is only four.
For an exact teaching calculation, consider the sharp null hypothesis that the layout assignment has no effect on any period's outcome. Also assume the design excludes carryover or other assignment-dependent changes between periods. Under that null, swapping AB to BA within a block reverses its observed difference while retaining its magnitude.
Four independent binary assignments give 2 to the fourth power, or 16, equally likely sign arrangements. The observed sum of differences is eight, the largest possible sum for magnitudes 2, 3, 1, and 2.
Only the all-positive arrangement reaches eight. A prespecified one-sided randomization test therefore has probability 1/16=0.0625 for a sum at least as large as observed. A two-sided test based on the absolute sum includes both all-positive and all-negative arrangements, giving 2/16=0.125.
The directional result looks consistent, but neither probability crosses a five-percent threshold. This does not prove the layouts are equivalent. It shows the limited resolution of this small randomized design and the distinction between an observed improvement and stronger inferential evidence.
If the study had forced exactly two AB and two BA blocks, there would instead be six permitted block-order assignments. The 16-arrangement calculation would not apply unchanged. An analysis must respect the actual assignment mechanism.
Suppose each period contained 20 individual tasks. There would be 160 task records across eight periods, but still only four randomized blocks. Treating all task records as independently randomized comparisons would misrepresent the design and could understate uncertainty.
Now consider a separate invented sequence in which the layout has no effect, but practice reduces completion time by one unit per period. The eight period averages are 20, 19, 18, 17, 16, 15, 14, and 13.
Using A for the first four periods and B for the last four would give A average 18.5 and B average 14.5. The apparent four-unit improvement is entirely explained by the time trend in this constructed example.
Under the earlier realized order AB, BA, BA, AB, A occupies periods one, four, six, and seven, with values 20, 17, 15, and 14. Their mean is 16.5. B occupies periods two, three, five, and eight, with values 19, 18, 16, and 13, also averaging 16.5.
This particular order cancels that particular linear trend. It does not guarantee protection against every trend or interaction. Randomization supplies a defensible comparison mechanism; it does not make all time-dependent complications disappear.
Persistent learning from one layout could affect later periods under the other layout. In that situation, merely alternating labels would not establish clean comparisons. The design and analysis would need to address the lasting effect rather than assume it away.

Illustrative case study

Fictional company Alder Interfaces runs the display-layout comparison with one consenting participant. It records period order, task difficulty, missing results, and the participant's experience.
The report presents the two-unit observed difference alongside the randomization analysis and design limitations. It does not advertise B as universally better for every user. This fictional case illustrates an individual evaluation and a reason to plan further evidence collection.

Common misconceptions and mistakes

One participant does not mean one measurement.
Many repeated records are not necessarily independent observations.
A before-and-after difference can reflect a time trend.
Carryover can undermine period comparisons.
Evidence for one individual does not automatically generalize to everyone.

Business use

Use individual experiments when personalization is the real question and repeated comparisons are feasible. In the fictional interface project, consider performance, usability, and measurement burden together. Keep the claim tied to the participant, intervention, and conditions actually studied.

Check your understanding

How many randomized blocks were used?

Answer: Four.

What is the observed mean A-minus-B difference?

Answer: Two.

What is the stated one-sided randomization probability?

Answer: 0.0625.

Related terms

Autocorrelation
Crossover trial
within-person comparison
randomization
carryover
period effect
personalized evaluation

Learn next

Compare blocked randomization with fixed alternation, then study carryover and analysis of repeated outcomes.

Sources

Office for Health Improvement and Disparities, N-of-1 study: comparative studies: https://www.gov.uk/guidance/n-of-1-study-comparative-studies . Interface trial, exact randomization calculation, trend example, and company scenario are original.

Previous: N-jet Next: N-pairs loss