How to read Phoenix research

There is no placebo group.

Nobody in Phoenix was given a dummy pill. When a member's number improves, some of that improvement can be expectation, attention, and the simple act of paying attention to your health. Placebo effects on subjective ratings like energy and mood are large and well documented. They are smaller, but not zero, on objective measures like ApoB and wearable sleep duration.

Nothing is randomised.

Members chose what to take and when to start. That choice is not random, and the reason for the choice can be the real cause of the result.

Confounding by indication.

People often start an intervention because a number is already moving. Someone starts berberine because their lipids are climbing. Someone starts a sleep protocol during a bad month. Both make the intervention look worse than it is. The reverse also happens: someone starts a supplement during a motivated, everything-is-going-well stretch, which makes it look better than it is.

Co-intervention, and it is severe here.

Phoenix members stack. Members with repeat bloodwork log an average of 12 concurrent interventions; members with no bloodwork log 0.4. So “has repeat bloodwork” and “is a heavy stacker” describe nearly the same people, and any single-intervention signal could be the rest of the stack talking. For every individual case we publish, we state how many other interventions that member started nearby. When that number is zero, the case is much stronger. We say which.

Self-selection.

Phoenix members are not a random sample of APOE4 carriers. They are people who found out their genotype, sought out a community, and paid to join it. 59% carry two copies of APOE4, against about 2.7% in the general population, about a 22-fold enrichment. That makes Phoenix an extraordinary place to study APOE4 and a poor place to estimate what is typical.

Survivorship.

The members with the longest data histories are the ones who stayed. People for whom nothing worked are more likely to have stopped logging, which quietly tilts every long-window result positive.

Daily check-ins are subjective.

Sleep quality, energy, mood, mental sharpness, calm and wellbeing are member-entered ratings on a 1–10 scale. They are real and they matter, but they are a person's impression of their day, not an instrument reading.

Some fields do not mean what their name says.

In the check-in data, mood and overall wellbeing are the same underlying field, so we never report them as two independent confirmations. The field named stress behaves as a calm score where higher is more relaxed. We publish these quirks rather than quietly working around them.

Lab noise can be bigger than the effect.

Every blood test has test-retest variability. HDL's noise floor is roughly ±7–8%, which is why we retracted a community-wide 4.6% HDL decline: the signal was smaller than the noise. Where a change sits inside the assay's own variability, we say so.

Units and naming are a minefield, and it has bitten us.

The same biomarker arrives from different labs under different names and in different units. ApoB and apolipoprotein_b are the same thing; LDL cholesterol, LDL particle number and LDL size are three different physical quantities that a careless match will average together into nonsense. Genotype arrives spelled four different ways. One naive match once discarded 866 of 1,277 members without any error appearing. We now canonicalise names, convert units explicitly, and print what we folded together.

Timing and lag are uncertain.

We know when a member logged a start. We do not always know when they actually started, whether they took it consistently, or at what dose. Dose is recorded on about 57% of entries and time of day is not captured at all. A follow-up blood test 77 days later is a real measurement at a real time, but the exposure between those two points is partly inferred.

Coverage is incomplete.

Not every member wears a device, retests on schedule, or logs every day. Missing data is rarely missing at random: people log more when things are going well.

Multiple comparisons.

Test a thousand things and roughly fifty will look significant by chance alone. Issue 01 tested around a thousand habit-by-outcome comparisons plus ten biomarkers. Where we report a corrected result, we say which correction. Where we report an individual case, we are explicitly not claiming it survived correction. It is one interesting person, presented as one interesting person.

We slice the group to find the signal, and we tell you which slice.

When we report a pattern, we have often looked at a subgroup: one genotype, one sex, an age band, the members whose starting number was worst. Choosing the cut after seeing the data makes a finding easier to produce by chance, and we are not going to pretend otherwise. Our answer is not to stop doing it, because a signal in a subgroup is exactly what a carrier wants to know about. Our answer is that the cut is always named in the sentence. If a number describes twelve women over 55, it says so. If it describes everyone, it says that too. Read the cut before you read the number.

Small numbers.

Many of these cohorts are single or double digits. An n=4 signal where all four move the same way is genuinely interesting and genuinely fragile. Both things are true.

Regression to the mean.

People often start something when a number is at its worst. The next reading tends to be closer to their average regardless of what they did. This alone can manufacture an impressive-looking improvement.

Underpowered is not the same as ineffective.

Our sharpest test of this: we ran statins → LDL through the pipeline as a positive control. Statin users dropped 18.5 mg/dL more than non-users, which is the right direction and the right clinical magnitude for the most certain drug effect in cardiology, and it still landed at p = 0.154. If that cannot reach significance at this sample size, then a null result here means we cannot see it yet, not that the thing does not work. We hold ourselves to that reading.

There is no adverse-event surveillance system.

Phoenix does not run structured safety monitoring. We cannot claim that no serious adverse events occurred among members, only that none were reported to us through the app.

Device and partner studies carry two extra limits.

Wearables are often not connected before a member's first use, so there is no clean group baseline to measure against. The comparison has to be use-nights against that member's own non-use nights instead. And when we report a device study, we report the discomfort too: in the ZenoWell taVNS study, alongside the positive sleep-onset reports, five members used pain language, two reported ear or skin irritation, two reported headache, and three used stop language. Positive experiences, null experiences and adverse experiences all get published.

This is not medical advice.

Phoenix does not diagnose, treat, cure or prevent any disease, including Alzheimer's. Medication decisions stay with your clinician. Phoenix does not check interactions with anything you already take. Nothing here is a recommendation to start, stop or change a treatment.