Your life is an n-of-1 study

The average person doesn't exist.

Population studies measure thousands of people to find what is true on average. The average is a summary of a crowd, and no single person matches it exactly. "Does sunlight improve mood?" is a question about the population. "Does sunlight improve this person's mood?" is a different question, and a harder one. Below is what the research says about weather, weekdays and mood, and why it points toward analysing a single life as an n-of-1 study: a study with one participant, run with the same care as a study with thousands.

The weather effect is real, small and individual

The largest careful study of weather and mood is by Denissen and colleagues, who linked the daily diaries of 1,233 people to local weather-station records and ran a multilevel analysis.¹ Temperature, wind and sunlight all had an effect on negative affect, but the average effect on mood was small in terms of explained variance.

Underneath that average, the variation between individuals was substantial, especially in the response to daylight. Some people are weather-sensitive, some are not, and a few respond in the opposite direction. The group mean blends all of them into one small number. A population study cannot tell which case applies to a given person. That person's own data can.

"Blue Monday" is mostly a memory effect

Mondays are widely believed to be the worst day of the week. Stone, Schneider and Harter examined day-of-week mood in roughly 340,000 people.² Weekends and Fridays did lift mood, but the Blue Monday effect was minimal: Monday was practically indistinguishable from Tuesday, Wednesday or Thursday.

The gap between memory and record is the interesting result. When participants were asked to recall their worst day, 65% named Monday, a day that was no worse than midweek in their own logged data. Memory does not store the week as it happened. It stores a story about the week, and the story is shaped by expectation. A check-in logged the same evening, before the story forms, is a much better measurement.

Why searching personal data for patterns goes wrong

If personal data beats both memory and population averages, mood apps should already be delivering reliable personal insights. Most do not, because searching one person's numbers for patterns runs into a well known problem in empirical science.

John Ioannidis described it in a widely cited 2005 paper, "Why Most Published Research Findings Are False."³ A result is more likely to be false when effect sizes are small, when samples are small, and when many relationships are tested with a lot of analytical flexibility. A personal tracker meets all three conditions: small effects, a sample of one person over a few months, and dozens of variable pairs as soon as ten things are tracked.

When enough pairs are tested, some of them reach significance by chance. That is a property of the arithmetic, and no amount of clever software removes it. The Spurious Correlations project illustrates it with series such as cheese consumption and deaths by bedsheet entanglement.⁴ An app that reports its strongest weekly correlation is mostly reporting noise.

The standard remedy is the Benjamini-Hochberg false-discovery-rate correction, used in genomics and clinical trials.⁵ It raises the threshold each result must clear in proportion to how many tests were run, which separates a finding that passed a fair test from the luckiest draw in a large set. One consequence of applying it is that an app reports nothing much more often. That is the correction working as intended.

N-of-1 trials are established science

Treating a single person as the whole sample is not unscientific. N-of-1 trials, multiple-crossover experiments run within one patient, are an established and growing method in medicine.⁶ When a condition is stable and responds reversibly to treatment, a well-run n-of-1 trial can match or exceed a conventional group trial for that individual, because it answers the question the individual has: what works for them.

A daily life tracker is, in spirit, a long observational n-of-1 study. It cannot randomise the weather or blind a person to their own sleep, so it establishes association, not causation. The unit of analysis is the same, though: one life.

What this means for Daynostic

These findings translate into a few design choices:

• Logging every evening, because memory reconstructs. The Blue Monday result is what happens when recall is trusted over records.²
• Sensitivity to small, individual effects. The weather literature shows the signal is small on average and specific to the person, so the analysis has to be able to find a small real effect without inventing one.¹
• Correction for the number of tests. Without false-discovery control, a tracker with many variables produces coincidences at a steady rate.³⁵
• Reporting only results that pass. The answer to "what is correlated this month?" is often "nothing yet". An analysis that cannot give that answer is not measuring anything.

The companion article on the analysis pipeline describes what an insight goes through before it appears: rank correlations, permutation tests against shuffled history, FDR correction and day-of-week controls. The research above is the reason that pipeline exists.