How to Read an N-of-1 Trial: A Randomized Experiment for One Person

An N-of-1 trial takes the machinery of a randomized trial and points it at one person: the participant takes treatment A during some periods and treatment B during others, in an order set by chance and usually concealed. What comes out is evidence about that person's response, and it holds only if the condition is stable, the treatment effects reverse between periods, the outcome is measured the same way throughout, and one period does not bleed into the next.

Begin with a real uncertainty#

The design earns its keep when nobody honestly knows which of two reasonable options suits the person in front of you; a population trial may show an average benefit while individual response varies. Or the evidence may simply be thin for someone with a rare condition or several coexisting illnesses.

An N-of-1 trial does not replace established urgent or life-saving treatment. It is also unnecessary when one option is clearly contraindicated, the expected effect is obvious and immediate, or the decision can be made from existing evidence and preference without a burdensome experiment.

The protocol should state the personal question before the first period begins. Is the comparison active treatment versus placebo, two active options, or a current regimen versus an alternative? What improvement would be large enough to change the decision? What adverse effect would outweigh that improvement? A vague goal invites a vague result.

The multiple-crossover structure#

A single switch from A to B is easily confounded by time, expectation, regression to the mean, or a naturally fluctuating symptom. N-of-1 trials repeat the contrast. A sequence might contain three pairs of treatment periods, with the order within each pair randomized.

Repeated periods serve two purposes. They show whether the same treatment tends to perform better more than once, and they estimate the natural variability of the outcome within the person; randomization prevents every A period from occurring earlier or during a predictably easier time. Blocking or balancing can keep the number of A and B periods similar.

The sequence should be generated before outcomes are known. If treatment order is changed whenever symptoms worsen, the trial becomes responsive care rather than a randomized experiment. Responsive care may be appropriate, but it answers a different question.

Choose a condition and treatment that can cross over#

The condition should persist through the study and remain stable enough that treatment periods are comparable. Chronic pain, sleep symptoms, attention, or recurring functional limitations can sometimes fit. Rapidly progressive illness, an unpredictable acute condition, or a problem likely to resolve during the trial usually does not.

Treatment effects should begin within a practical period and reverse after stopping, and a curative procedure, vaccine, durable behavioral training, or intervention that permanently changes disease course cannot be reset for the next period. Treatments that are unsafe to stop or restart are unsuitable.

Stability is a matter of degree. Symptoms may fluctuate, which is one reason to repeat periods. The design fails when the underlying trajectory overwhelms the treatment contrast or when an outside event changes the condition irreversibly halfway through the sequence.

Period length and washout are biological decisions#

Each period must be long enough for the treatment to take effect and for the chosen outcome to be measured reliably. A period that is too short can miss benefit. One that is unnecessarily long adds burden and increases the chance that background changes complicate the trial.

Washout prevents the prior treatment from affecting the next period. It should account for both drug clearance and the duration of downstream effects. Several half-lives may reduce the amount of drug while symptoms, physiology, or withdrawal continue to reflect the prior period.

Some protocols include a wash-in phase and analyze only a stable portion of each period. Others use outcomes that respond rapidly enough that no separate gap is needed; the report should justify the timing from the intervention's expected onset and offset rather than leave you with a generic interval.

Carryover cannot be repaired reliably by a weak post hoc test. If it is substantial, the later period no longer represents a clean contrast. Sensitivity analyses that omit early measurements can help, but design remains the main protection.

Blinding separates treatment from expectation#

Symptoms can change because a person expects one treatment to help, recognizes a side effect, or knows the calendar. Whenever practical, treatments can be prepared to look and taste alike, with the participant, clinician, and outcome assessor unaware of period assignment until the sequence is complete.

Blinding is not always possible. A device, exercise program, or recognizable adverse effect may reveal the assignment. A nonblinded trial can still provide information, but subjective outcomes are then more vulnerable to expectation. Objective measures do not remove every bias, since adherence and behavior can change when assignment is known.

Reports should say who was blinded, how treatments were masked, when the code was broken, and whether the participant guessed assignments. “Double blind” is less informative than a direct description.

Outcomes should be personal and measurable#

The primary outcome should matter to the decision. A daily symptom score, walking distance, sleep measure, rescue-medication use, or functional task may be more informative than a biomarker with no clear personal meaning. Harms and treatment burden belong beside benefit.

Measurement must be frequent enough to summarize each period and consistent enough for comparison. The protocol should specify the scale, timing, data source, and period summary before seeing results, and choosing whichever outcome looks most favorable after the trial creates the same selective-reporting problem found in larger studies.

A minimally important difference helps interpretation. A statistically detectable average difference may be too small to notice or worth less than the inconvenience and adverse effects, and conversely, a consistent meaningful improvement can guide care even when a formal test has limited power.

Adherence should be measured when feasible. A treatment cannot be judged fairly from a period in which it was rarely used, but excluding low-adherence periods after seeing outcomes can bias the answer, and the estimand should clarify whether the question concerns assignment to a regimen or response while following it.

Analysis should show the pattern, not only one number#

A time plot of outcomes by day, treatment, and period often reveals information a single average conceals. It can show delayed onset, carryover, a time trend, outliers, or a difference that appears only in one cycle.

Formal analysis can compare period means, use regression that accounts for time and serial correlation, or apply Bayesian methods that combine prior evidence with the participant's data. The choice should be prespecified and proportionate to the amount and quality of information.

Measurements on adjacent days are correlated, and dozens of daily scores are not dozens of unrelated participants. An analysis that treats each observation as separate will overstate precision, so the interval you read will be narrower than the evidence supports. Period and cycle structure should remain visible.

Reading the result well asks you to hold several things at once: how big the difference was, whether it repeated across cycles, how uncertain it is, what it cost in harms, whether the participant actually took the treatment, and which option they would rather live with. A binary declaration that treatment A “won” can misrepresent a close or heterogeneous pattern.

Missing data and early stopping can bias the result#

Daily recording is burdensome. Missing outcomes may cluster during severe symptoms or inconvenient treatments, making them informative rather than random. A participant may also stop a period because of harm or obvious lack of benefit. Those events are part of the result, not mere data-cleaning problems.

The protocol should define rescue treatment, stopping criteria, handling of missed doses, and what happens after an adverse event. Participant safety takes priority over completing a balanced sequence. The report should then explain how any early stop limits the planned comparison.

An interim rule can allow stopping when one treatment is clearly superior or unacceptable, but repeatedly looking at the data and stopping informally can exaggerate the apparent difference. Prespecification makes the tradeoff between efficiency and bias visible.

Individual value and population limits#

A completed N-of-1 trial can provide unusually direct evidence for the participant studied. It estimates response in that context, with those coexisting conditions, preferences, outcomes, and treatment periods. This local relevance is its main strength.

The same specificity limits generalization. One person's result does not establish an average effect for others. A prospectively coordinated series of N-of-1 trials can pool person-specific estimates and study response variation, but the protocol and analysis must account for both within-person and between-person structure.

The result also may not predict long-term benefit or rare harm if periods were short. Population trials, pharmacovigilance, mechanistic evidence, and the N-of-1 result answer complementary questions.

What the report should include#

The CENT reporting guideline extends CONSORT for individual and series N-of-1 trials, alongside the general CONSORT framework updated in 2025. Read the paper looking for the design first: the rationale, who the participant was, what the interventions were, how the sequence was generated, how allocation was concealed, and who was blinded. Then look for the mechanics: period and washout timing, the outcomes, the analytic plan, the harms, and a complete flow of every period.

The report should identify deviations from the protocol and show raw or summarized outcome trajectories, and it should distinguish a single trial intended for one decision from a planned series intended to support broader inference.