The Gold Standard of Experimental Design | N-of-1 Trials, Individual Treatment Effects and Repeated Randomisation

1. Ask whether the decision is individual and repeatable

Back to contents

An N-of-1 trial fits a decision about which of two plausible treatments works better for one person when the condition is reasonably stable, effects begin and recede on a useful time scale and outcomes can be measured repeatedly. It is poorly suited to curative treatment, irreversible harm, vaccination, surgery or rapidly progressive disease. Define the decision first: symptom relief, function, adverse effects or a weighted combination. The aim is an individual treatment effect for a specified period and person.

2. Build repeated treatment periods, not a casual alternation

Back to contents

A credible design repeats each treatment in several periods so that one unusually good week does not decide the result. Period length must allow the treatment to act and enough measurements to stabilise. Sequence length must remain feasible for the participant. ABAB is easy to understand but predictable and confounded with any repeating time pattern; randomly selected balanced sequences provide stronger protection. Replication within the person is the effective evidence base.

3. Randomise the sequence and conceal it when possible

Back to contents

Randomisation breaks systematic links between treatment and unmeasured time-varying conditions in expectation. Balanced blocks can ensure each treatment appears equally often while preserving unpredictability. Blinding may reduce expectation and reporting effects when matched treatments are feasible, but it is not always possible for behavioural or device interventions. Store the allocation sequence, concealment process and deviations. A schedule chosen after seeing symptoms is adaptive care, not the same experiment.

4. Measure outcomes at the right frequency

Back to contents

Choose an outcome meaningful to the participant and responsive to treatment. Record it often enough to observe within-period change without creating unsustainable burden. Wearables can add dense data but also missingness, calibration drift and autocorrelation. Predefine the primary summary, such as average pain in the final three days of each period, and retain adverse effects and adherence. More observations do not compensate for measuring the wrong construct.

5. Separate treatment response from time and carryover

Back to contents

Symptoms may trend, cycle by weekday or respond slowly after treatment changes. A previous treatment can influence the next period. Use pharmacology or mechanism to choose run-in and washout rules, and exclude or model transition observations when justified. Do not rely on a low-powered test to prove no carryover. The related crossover design principles remain important, but an N-of-1 trial additionally asks whether the result supports this person’s decision rather than a population average alone.

6. Worked example: two migraine preventives

Back to contents

One person compares treatments A and B in six randomised four-week periods. The primary outcome is migraine days during weeks three and four, allowing two weeks for onset. Counts are 4, 7 and 5 under A, and 8, 9 and 6 under B. The paired period contrast favours A by about three days, but the last B period coincides with exam stress and missing sleep data. A prespecified time-adjusted model still favours A, though with wide uncertainty. The decision combines effect size, adverse effects and the participant’s minimum worthwhile difference rather than declaring a universal winner.

7. Model serial correlation and period effects

Back to contents

Daily outcomes from one person are correlated; yesterday’s symptom can predict today’s. Treating every day as independent makes precision look too high. Analysis can summarise periods or use time-series, mixed or Bayesian models that represent autocorrelation, trend, seasonality and carryover. With few periods, complex models are weakly identified. Show raw time plots and period summaries alongside the model so that the inference is not carried by invisible assumptions.

8. Changed-context transfer check: study timing

Back to contents

A student wants to compare studying mathematics before school with studying after dinner. Weeks are randomised, each schedule is repeated and a short mixed problem set plus fatigue rating is collected. School topics change over time and practice improves performance, so identical test forms would add learning and different forms need calibration. The transferable design uses repeated randomisation and personal outcomes, but it cannot blind timing and must model week and topic. It supports that student’s routine, not a claim about every student.

9. Know the limits and common misconceptions

Back to contents

N-of-1 does not mean one observation, and repeated measurement without randomised treatment periods is not automatically a trial. A clear individual result may not generalise to others. Regression to the mean, co-interventions, nonadherence and selective recording can mimic benefit. Some effects are too delayed or persistent for reversal, while severe instability makes periods incomparable. Ethical monitoring and clinical oversight remain necessary; individualisation is not permission to experiment without safeguards.

10. Audit allocation, adherence and decision rules

Back to contents

Archive the protocol, eligibility rationale, sequence generation, concealment, period and transition rules, outcome instruments, timestamps, adherence, concurrent changes and adverse events. Plot the full series with treatment periods marked. Reanalyse plausible trend, autocorrelation and missing-data choices, and show how the conclusion changes under different minimum worthwhile effects. Record the final shared decision and why the evidence was or was not sufficient.

11. Frequently asked questions

Back to contents

Is one person enough? Enough for an individual decision when repeated randomised periods are informative, not for every population claim. Must the trial be blinded? Blinding helps but is not always feasible. How many periods are needed? Enough replication for the expected variability and effect, balanced against burden. Can results be combined across people? Yes, aggregated N-of-1 analyses can estimate population patterns while retaining individual effects. Is self-tracking equivalent? No; treatment assignment and prespecification matter.

12. The gold standard turns personal care into a visible experiment

Back to contents

A strong N-of-1 trial aligns one person’s decision with repeated, randomised and interpretable treatment periods. It measures outcomes that matter, respects onset and persistence, represents time dependence and reports raw trajectories with uncertainty. The gold standard neither overgeneralises a personal result nor dismisses it for lacking a group average. It makes the individual causal comparison explicit enough to support a careful decision.