baselineMethod write-ups, full records from the cohort, and what the benchmark turns up. We publish the analyses that did not settle alongside the ones that did.
Among 547 men taking no lipid medication, mean total cholesterol in the top baseline fifth fell from 253 to 237 mg/dl over a year, and in the bottom fifth rose from 161 to 165. Nobody was treated. That drift is what your supplement has to beat.
In a pooled analysis of 52 trials, perceived recovery improved by 0.66 standard deviations and strength recovery by 0.12 with an interval running from −0.38 to 0.62. Same trials, same timepoint, opposite conclusions.
Across 37 three-armed trials, the no-treatment arm improved by an SMD of −0.24 and the placebo arm by −0.44. An uncontrolled before-and-after on your own body returns the sum of both, plus whatever the intervention did.
Across 536 nights against a chest-strap ECG, five wearables estimated nocturnal RMSSD to within 5 ms — and still differed significantly from each other, because they compute a different metric over a different slice of the night from a signal that is not the heartbeat.
The direct evidence is one twelve-person pilot that reported no change and printed no numbers. Everything else people cite is post-exercise heart rate, a side-effect tally, or a cardiac literature that never listed heart rate as an outcome.
In a published power analysis, 400 measurements split into forty short alternating periods detected an effect at 0.681 power where the same 400 measurements in two long blocks managed 0.126. Nothing was added — the order changed.
Three randomised trials, 151 people, and a headline result that does not survive contact with the second outcome. Why magnesium is our worked example — and why that is not an endorsement.
Across six wrist wearables scored against polysomnography, deep-sleep estimates ran from 25 minutes short to 44 minutes long. The effects people chase are smaller than that. Here is why within-person tracking still works.
A p-value does not tell you how large an effect is, or how sure you should be. The American Statistical Association said so plainly in 2016. Here is what we report instead, and what it costs us.
Your own 62 nights are not enough on their own. The population average is not about you. Hierarchical models sit in between — and the literature is specific about where they break.
Every system we tested infers causation from correlation somewhere. Here is where each one breaks, and how we score calibration.