N-of-1 Inference

N-of-1 inference is statistical reasoning about one individual, where the replicates are repeated measurements on that individual over time instead of different people. A population trial asks whether a treatment shifts the mean in a group. An N-of-1 analysis asks whether something changes you, and it gets its power from measuring you many times under known conditions.

The phase 1/2/3 framing people search for is the population machinery: phase 1 is roughly 20-100 participants and looks at safety and dose range, phase 2 is a few hundred and looks for a signal, phase 3 is typically several hundred to several thousand and is sized to detect a specific effect size at a chosen power. N-of-1 designs sit orthogonal to that ladder. The classic form is a multi-crossover trial: alternating A/B periods within one person, randomized in order, ideally blinded, with a washout between periods. Aggregated across several individuals, these become “series of N-of-1 trials” and can estimate both individual and average effects.

How it works

Three ingredients make single-subject data inferential rather than anecdotal.

Repetition. You need multiple periods per condition. Two periods (one A, one B) gives you no way to separate the condition from the trend. Four to six crossover pairs is a common target.

Randomization. If you randomize the order of periods yourself, the randomization distribution is a valid null: permute the assignment labels across all orderings that could have occurred, recompute your test statistic, and read off the p-value. This is exact and makes no distributional assumption. Recent theory works out what randomization inference can and cannot recover when N equals one, and the answer is that it identifies effects for the sequence you ran, under assumptions about carryover ¹.

Control of the time axis. Within-person series are autocorrelated. Fasting glucose today predicts fasting glucose tomorrow. A t-test on daily values treats 60 days as 60 independent observations when the effective sample size may be closer to 15. Fit an AR(1) or use block permutation.

The molecular version of this is younger and less standardized. Single-subject transcriptomics compares two states of the same person and scores pathway-level deviation instead of gene-level fold change, which stabilizes the estimate when you have one sample per state ². Clinical N-of-1 programs in rare disease and oncology now use individual molecular profiles to pick an intervention for one patient, with the profile itself as the design input ³⁴.

In your own data

Where it shows up in files you will have:

• CGM: a CSV with a timestamp column and a glucose column at 1- or 5-minute resolution. This is the densest N-of-1 substrate you own. A standardized meal repeated on randomized days, with iAUC computed over a fixed 0-120 min window and trapezoidal integration above baseline, gives you a real within-person estimate. Check that your baseline is a stable pre-meal 15-minute mean, not a single reading.
• RNA-seq: a counts matrix (gene_id × sample) from salmon quant or featureCounts. With one sample per condition, DESeq2 cannot estimate dispersion from replicates. Either supply dispersion manually via estimateDispersionsGeneEst with a fixed prior, or move to a pathway-distance statistic across paired samples ². Do not run results() on a two-sample design and report the p-values as though they mean something.
• Proteomics and blood panels: long-format tables with an analyte, a value, and a draw date. The number to know per analyte is intra-individual biological variation plus assay CV. If your CRP CV is 40% within-person, a 30% change between two draws is noise.

Common mistakes: treating a single pre/post pair as evidence; ignoring the order effect when you never randomized; and running the analysis after looking at the data, which converts an exact test into a fishing expedition. Write the analysis script and the primary endpoint before you collect period 1.

Limitations

N-of-1 results generalize to you under the conditions you tested and nothing more. Carryover between periods biases the estimate toward null if washout is too short. Conditions that resolve or progress on their own are poor candidates, because the trend confounds the assignment. Aggregation frameworks and reporting standards for single-subject research exist but are unevenly applied ⁵⁶, and where these designs enter therapeutic development they require regulatory and clinical infrastructure that an individual does not have ⁷⁸. Anything touching a diagnosis, a drug, or a dose belongs with a clinician, and rare-disease or oncology work of this kind is run under clinical protocols ⁹.

What single-subject analysis is good for on your own: detecting effects that population averages hide, which wearable-data studies have shown can be large and individual-specific ¹⁰.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.