Correlation metadata is the structured record Unfair keeps of how dose events line up in time with outcome observations, expressed as a family of statistics with explicit lag, direction, and confidence. Every paired dose-outcome series in the app carries a Pearson r, a Spearman ρ, a best-fit lag in days, and a noise floor estimated from the pre-intervention baseline. These numbers are not claims of causation; they are the raw material the recommendation ranking reads when it decides whether an ingredient deserves more or less weight in the next cycle.
What correlation metadata contains
Four statistics ship with every tracked series, and each answers a different question about the dose-outcome pair.
| Statistic | What it answers | Useful when |
|---|---|---|
| Pearson r | How tightly a linear relationship fits the two series | Both series are roughly continuous and symmetric (HRV, sleep efficiency) |
| Spearman ρ | How tightly a monotonic relationship fits, without assuming linearity | One series is ordinal or skewed (1–5 focus score, joint pain VAS) |
| Best-fit lag (days) | At which day-offset the relationship is strongest | Compounds with overnight-to-multi-day action (magnesium, adaptogens) |
| Baseline noise floor | The r a random pairing of baseline days would produce on average | Deciding whether an observed r is distinguishable from chance |
The noise floor matters more than users expect. For a 28-day log with typical day-to-day variance, random pairings commonly produce |r| in the 0.12–0.18 range, so a headline correlation of 0.20 is barely above noise even though it looks non-zero.
A worked lag-correlation example
A user runs a 45-day magnesium glycinate protocol at 400 mg, 60 minutes pre-bed, while syncing Oura sleep. Unfair computes Pearson r between dose-day and several sleep outcomes at day-offsets 0 through 3:
| Outcome | r at lag 0 | r at lag 1 | r at lag 2 | r at lag 3 |
|---|---|---|---|---|
| Sleep onset latency (min) | -0.12 | -0.28 | -0.15 | -0.06 |
| Sleep efficiency (%) | 0.09 | 0.34 | 0.21 | 0.08 |
| Deep sleep (min) | 0.07 | 0.19 | 0.26 | 0.11 |
The strongest pairing is magnesium on day 0 with sleep efficiency on day 1, at r = 0.34. The baseline noise floor for this user's log density is 0.16, so 0.34 is roughly 2.1× the floor — a usable signal. Deep sleep peaks a day later at lag 2, which matches the common user report that glycinate produces a noticeable sleep-architecture shift on the second night.
None of this proves the magnesium caused the shift. A n-of-1 experiment with an on-off structure is the right follow-up when the correlation clears the noise floor and the user wants a firmer answer than temporal co-occurrence.
What correlation metadata does not claim
Unfair draws a hard line between three questions and keeps them visually separate in the review screen:
- "Did these events co-occur?" — yes/no, answered by timestamp alignment.
- "How tightly do they move together?" — answered by Pearson r and Spearman ρ.
- "Did the dose cause the outcome?" — not answered by any single correlation; requires repetition, withdrawal-return, or a controlled trial.
Correlation values above 0.5 in self-tracked supplement data are unusual and almost always reflect a shared confounder (a training block, a sleep-schedule change, an illness window) rather than a direct effect. The engine treats anything above 0.6 as a flag to investigate confounds before updating rank, not as a green light.
Limits and failure modes
Four failure modes show up repeatedly in user logs:
- Irregular logging creates artificial lag patterns because the dose timestamp drifts later in the day, not because the compound moves slower. Fast entry paths matter here for the same reason they matter for adherence — late timestamps poison lag estimation.
- Short series (under 21 days) produce unstable r values. A single bad night can shift a 14-day r by 0.2 or more.
- Missing days default to "no value"; the system never fills gaps silently. A log with under 60% coverage is flagged and down-weighted rather than computed and displayed.
- Self-report outcomes are vulnerable to placebo expectancy in the first weeks of a new compound, which can inflate early r values and produce retractions once the expectancy window closes.
How this appears in Unfair
Each tracked dose-outcome pair shows the four statistics with the best-fit lag highlighted, the noise floor printed as a ghost line, and a plain-language summary ("magnesium correlates with higher sleep efficiency the next day at r = 0.34 across 45 days, clearing the noise floor"). These correlations feed the weighted ranking, but never directly; they pass through confidence bands, adherence gates, and repetition checks first. This is one of the advanced capabilities that separates Unfair's review surface from apps that simply chart doses over outcomes without a statistical layer.
Clinical safety note
Correlation metadata is for self-experimentation within a personal baseline, not for medical decisions. A correlation between a supplement and a worrying symptom — rising resting heart rate, new palpitations, mood drop — is a reason to pause the stack and involve a clinician, not a reason to keep logging for a tighter r value.