When the numbers change, we ask what moved them.
The signals look different early versus late. The tempting move is to call that a story: the body changed. The honest move is to list everything that could have done it, find out which ones the data can actually speak to, and subtract the ones we pin down before reading the rest.
What else could explain this?
When the numbers look different early versus late, the honest question isn't 'what does it mean'; it's 'what could be moving them.' A recovering body, better pacing, a medication, the seasons, or plain chance could each do it, and with one person and 29 crashes we usually can't tell which. So we don't pretend to. We do something better: we name every candidate, investigate the ones the data can speak to, and when we pin one down, we subtract it before reading the rest.
- 1Don't attribute the change to a single cause we can't separate, a difference over time is a number, not a story.
- 2Do qualify, run the research the design allows: is a driver even visible, and if so, how big?
- 3Do feed back, a driver shown real becomes a correction we model out, so every other signal is read net of it.
- Visible, not yet bounded An effect shows up, but we can't put a number on its size yet.
- Modelled out Established and now corrected for, subtracted before we read the signals it affects.
- Irreducible Can't be removed, only acknowledged, the floor of a one-person study.
- Bounded Characterised, size, direction, and which signals it touches.
- Looked, not visible We tested for it and found no detectable effect within our power.
| Driver | Stress | Overnight stress (HRV proxy) | Resting heart rate | Body battery | Exertion / load | Crash frequency & depth |
|---|---|---|---|---|---|---|
| Recovery Visible, not yet bounded | ||||||
| Pacing practice Visible, not yet bounded | ||||||
| Citalopram Modelled out | ||||||
| CPAP Irreducible | ||||||
| Season Visible, not yet bounded | ||||||
| Fitness history Bounded | ||||||
| Weight Bounded | ||||||
| Aging Bounded | ||||||
| Chance Looked, not visible | ||||||
| Measurement regime Looked, not visible |
Real-world drivers
Things that genuinely changed the body or the signal.
-
Recovery
Visible, not yet boundedThe body genuinely getting better, Long COVID recovery across the four years.
Could touch Crash frequency & depthResting heart rateOvernight stress (HRV proxy)
Crashes did get rarer and milder over the four years; that change is real and described. And the lived recovery-phase boundaries I drew from memory, never tuned to the watch, mostly show up independently in the data: five of the six surface as real multi-channel shifts (one boundary falls outside the recorded window). But how much of the over-time change is recovery itself, versus better pacing, the medication, or chance, the design can't separate on one person.
Visible at the boundaries but not decomposed into a per-day magnitude, and it runs underneath the same years the medication taper did, so 'recovery' can't be cleanly isolated from the other drivers that moved alongside it.
-
Pacing practice
Visible, not yet boundedGetting better at pacing, the very skill Wiggers' guide teaches. Crucially, the watch was used live to pace, so here it was a lever, not just a gauge.
Could touch Exertion / loadStressCrash frequency & depth
I pace off the watch and my felt state every day (five channels I act on live) so pacing isn't an outside force on an untouched body; it's woven into the very data these tests read. Fidelity improved over the years (partial early, steady only recently). Because good pacing prevents crashes, a quiet signal can mean a crash was averted rather than never coming, so pacing confounds the trust metrics too, not only the time trend.
The instrument was part of the intervention, and whether the pacing actually works is a separate question we haven't measured; this is the hardest driver to disentangle.
-
Citalopram
Modelled outAn SSRI taken across these years, with a dose that changed over time.
Could touch StressOvernight stress (HRV proxy)Body battery
Confirmed, dose-graded effect on three channels: more drug raised measured stress and lowered the body-battery floor. (Why is a separate question, the plausible-but-unproven reason is that SSRIs can blunt heart-rate variability, which the watch reads as higher stress; a literature review found that mechanism mixed, not established, so we hold the measurement, not the explanation.) Because the dose is known day by day, we could finally do what we promised: model it out, subtract the dose effect and re-read the signals net of the drug. The honest result is undramatic. Most of the scorecard's signals are built from night-to-night *changes*, and a slowly-moving dose largely cancels out of a change, so the correction barely moves them, and none got stronger once the drug was subtracted. And the dose effect itself held up under a confounder re-audit: it survives weight gain, deconditioning, aging, and season, the drug's fingerprint is real, not those slow trends in disguise (the betas point the same way whether the dose is going up or down, which a seasonal or steadily-rising confound could not fake).
Now modelled out on the three confirmed channels; the correction is small because the change-based signals were never strongly dose-driven. On resting heart rate it cuts the other way, the drug lowers the number by about a beat at the top dose, which masks part of the illness-driven rise there. One honest scope note: subtracting the dose cleans these channels of the drug, not of everything, a long-range, across-the-years read would still owe weight, the deconditioning tail, and aging their own corrections, which is part of why the crash-precursor tests are kept to short windows. Everything we know and tested is pulled together under 'The citalopram question'.
-
CPAP
IrreducibleA CPAP trial for sleep-disordered breathing, which ended in spring 2024.
Could touch Overnight stress (HRV proxy)Body battery
The CPAP trial ended just seven days before citalopram began. That collision makes the whole spring-2024 window structurally unanalysable (every before-and-after comparison is either empty or shared with the drug) so CPAP's own effect on the overnight channels can't be separated out by any analysis this data allows.
Not a gap we can close: the two interventions sit too close in time. Acknowledged, not removed, the floor of a one-person record.
-
Season
Visible, not yet boundedWinter versus summer, ambient temperature and daylight move resting heart rate, sleep, and activity.
Could touch Resting heart rateOvernight stress (HRV proxy)Exertion / load
A known confounder for any across-time comparison. The pre-illness baseline is only about seven months (a single winter) so season and illness-state are entangled there by construction. One reassurance from inside the medication taper: a flat spring-2025 stretch, with the dose held steady, showed nothing moving, which argues against a simple 'spring made it better' story. A proper season-stratified decomposition is specified but runs only if we ever make a pre-illness-vs-Long-COVID baseline comparison, which the scorecard avoids by staying inside the illness era.
Visible as a structural confound, not decomposed into a number, and not scheduled, because nothing currently on the site needs that pre-illness baseline contrast.
-
Fitness history
BoundedA lot of endurance sport before COVID, so my resting heart rate was probably trained low, and would drift back up as activity fell away with the illness.
Could touch Resting heart rate
Examined now, and fitness turns out to be a real but minor player, not the engine. My resting heart rate did rise across the four years, but that rise is NOT mostly lost fitness: deconditioning is front-loaded, finishing and plateauing inside the first year, while my resting heart rate kept climbing into year four. So the later rise isn't explained by lost fitness. At the other end of the timeline the same driver cuts the other way: right before COVID I was at my athletic resting-heart-rate floor, and that peak fitness masked the infection's heart-rate rise in the COVID-window check, which is why that result was carried by the stress and body-battery channels, not resting heart rate.
Characterised as a minor, front-loaded contributor, not the driver of the multi-year rise. One honesty note: my fitness decline after early 2022 is a modelled estimate, not measured (my last real recorded run was March 2022, so the watch's later fitness number is its own decay guess).
-
Weight
BoundedBody weight climbed a lot over these years, from an athlete's ~74 kg to about 89 kg now.
Could touch Resting heart rate
A real, slow driver of resting heart rate, and now measured. Across 56 weigh-ins the effect works out to about 0.32 bpm per kilogram, in line with the general literature, so weight gain accounts for part of the multi-year resting-heart-rate rise. It's sparsely sampled, though: a roughly 14-month gap in weigh-ins sits right over the illness onset, so its exact share stays fuzzy.
Measured and bounded, but thinly sampled over the key onset window.
-
Aging
BoundedFour years older, age alone nudges resting heart rate a little.
Could touch Resting heart rate
Bounded, and minor. In this age band (early-to-mid forties), resting-heart-rate change from aging is somewhere between nothing and about 0.3 bpm a year, far too small to explain the trend. Named so it's accounted for, not because it carries any weight.
A small, bounded contributor, ruled in for completeness, ruled out as an explanation.
Artifact drivers
Things that change the apparent pattern without anything in the body changing, not every difference over time is even about the body.
-
Chance
Looked, not visiblePlain luck. With 29 crashes in one person, the numbers wobble, and testing many signals at once gives chance more ways to fool you.
Could touch StressOvernight stress (HRV proxy)Resting heart rateBody batteryExertion / loadCrash frequency & depth
We partly accounted for luck. The seven tested channels aren't seven independent witnesses; they collapse to only about three or four, so the honest significance bar is stricter than it first looks. After that correction, no over-time story survives as more than wobble. (A full, corpus-wide false-discovery count across every test we ran isn't assembled; that piece stays un-examined.)
The wide error at 29 crashes (confidence intervals spanning thirty points or more) is the irreducible floor of a one-person study. We report it rather than hide it.
-
Measurement regime
Looked, not visibleWhat the watch could record changed over time, some signals only exist in the recent years.
Could touch Overnight stress (HRV proxy)Body battery
We checked, and for the tested channels the answer is clean: all seven sit in the watch's full-coverage daily data, present across the whole record. So an over-time difference in a tested channel is a change in the body or behaviour, not an artifact of when the watch started recording it. The availability artifact is real, but it's confined to other, sharper channels we don't test on, like per-minute overnight body-battery, which only fills in from about mid-2024.
Clean for the tested channels; a live risk only for the finer untested channels, whose coverage windows are known and flagged before any over-time reading.