Loading Longvai…
Longvai is built on conservative, well-understood statistics — not on hype. This is what we run, why we picked it, and where we draw the line on what we'll claim.
Longvai is a wellness product, not a medical device. If your data drifts somewhere worrying, we'll flag it, give you the timeline, and hand you a clean PDF for your doctor. Diagnosis is your physician's job. Pattern recognition is ours.
Every claim is labeled. CAUSAL only for things you A/B-tested on yourself. STRONG CORRELATION for high-confidence associations. SUGGESTIVE for early signal. We don't blur these.
Sample too small? Confidence too low? Confound too plausible? We say so. A confident “wait a few weeks” beats a confident wrong answer, every time.
No streaks. No daily scores designed to make you anxious. No notifications engineered to pull you back. The point is to read it on Sunday and live your week.
No black boxes. Every method below has a paper or a textbook chapter behind it — and a reason we use this one and not the obvious alternative.
Used for: Ranking the candidate drivers of any outcome (sleep, HRV, glucose, etc.)
Spearman beats Pearson here because your data is messy — outliers from a hangover or a sick day would otherwise dominate a Pearson coefficient. Rank correlation is robust by design.
Used for: Experiment readouts — is the B-condition median actually different from A?
Non-parametric, doesn't assume your sleep distribution is normal (it isn't), works with small samples (n=14 nights), and resists outliers. The default test for every n=1 trial.
Used for: Telling you whether a significant finding is also a useful one.
A p-value tells you the effect probably isn't zero. d tells you how big it is. d = 0.2 is small, 0.5 medium, 0.8+ large. We surface d on every comparison — because a p=0.001 with d=0.05 is statistically real and practically meaningless.
Used for: Detecting when something in your data actually changed.
If your deep sleep drops, you want to know whether the change started on April 14th or has been slowly drifting. Bayesian methods give a posterior over change-points with calibrated uncertainty — no arbitrary thresholds.
Used for: Anomaly detection on your individual baselines.
Adapted from manufacturing process control. We compute your personal mean and variance for each biomarker over a rolling window, then flag deviations beyond k = 2–3 standard deviations. The window adapts to cycle phase, travel, and training load.
Used for: Telling you when an experiment has enough data to conclude.
Before you start a trial, Longvai estimates the sample size needed to detect a meaningful effect with 80% power. During the trial, it shows you running power — so you know when to stop, and don't fall into the 'peeking' trap that fakes significance.
For every biomarker, we keep two ranges side-by-side — the population reference from peer-reviewed sources, and the personal range we computed from your timeline. You see both, always.
End-to-end encrypted at rest and in transit. Your health data never trains any shared model without explicit opt-in. You can export everything — in CSV and FHIR format — in one click, anytime, even if you cancel.
AES-256 at rest. TLS 1.3 in transit. Encryption keys you own, not us.
Pattern analysis runs on your device for the highest-sensitivity correlations. Nothing leaves.
Your data never improves someone else's model unless you opt in to the anonymized community dataset.
Your full health timeline in the standard medical interoperability format. One click, immediate.
No ads. No data brokerage. No affiliate deals on what we recommend.