How many nights does creatine need before you can judge it (measured on sleep latency)?

Twenty-eight nights. The effect size documented in the literature — 50.82% against the personal baseline — clears the 28-night detection threshold for sleep latency in our dataset (45.38%). At that window, you have enough personal context to judge. What the only double-blind placebo-controlled trial with objective actigraphy actually found: creatine made sleep latency 5.59 minutes worse than placebo, with no statistical significance (t=0.98; p=0.35). The cleanest population in the corpus. The clearest null in the dataset.

What the study measured

Ben Maaoui et al. (2025) published in Nutrients (17(24):3831) the first crossover double-blind placebo-controlled trial using objective actigraphy to evaluate creatine and sleep. The sample was deliberately narrow: 14 healthy, active men with no sleep disorders (PSQI below 5), no medications, no supplements. Subjects who already slept well, measured by device rather than self-report.

If creatine does anything to sleep onset time, this is where it would appear. The setup removed every confound it could.

Dose: 20 g per day for 7 days — the loading protocol used to saturate muscle creatine stores rapidly. Most users take 3–5 g per day on an ongoing basis. The study tested the higher-load phase.

The numbers: sleep latency with creatine, 30.68 ± 14.59 minutes. With placebo, 25.09 ± 20.22 minutes. Creatine was 5.59 minutes slower. t=0.98, p=0.35. No statistical significance. No objective effect.

What did move: subjective sleep quality. Cohen's d was 0.81, p was 0.009. Participants reported sleeping better. The actigraphy recorded no improvement in onset latency. That gap — felt improvement with no objective signal — appears across multiple compounds in the corpus. It is not a minor distinction. It is evidence that internal sleep quality perception is not a reliable measurement instrument.

How "when to judge" gets calculated

The study answers what happened in 14 people. The house answers when you can judge it in your own body. These are different questions, and they require different tools.

The house dataset holds 650 nights of one body. Sleep latency in that body is extremely noisy: coefficient of variation 84.41%, autocorrelation 0.026 (near-pure noise), relative coefficient of variation 233.96%. Night to night, this metric does not maintain a stable baseline. Noise governs it more than any underlying trend.

To detect a real signal against that noise, the house calculates a relative change threshold from its own nights. At 28 nights, the threshold for sleep latency is 45.38%. That means: a change must represent at least 45.38% of the personal median (11 minutes) before it rises above statistical noise into signal territory.

The effect documented in the study — 50.82% of the personal median, expressed as the 5.59-minute difference over a base of 11 — clears that threshold. That is why the answer is 28 nights: an effect of that magnitude would, in principle, be detectable at that window with this instrument.

Read the number with its direction. 50.82% above the noise floor — but the effect went the wrong way. Latency lengthened, not shortened.

Detection threshold by follow-up window

Nights of follow-up	Detectable change threshold
28	45.38%
56	32.09%
84	26.2%
112	22.69%
140	20.29%
168	18.53%

Source: ruido.medido.json#sleep_latency_min · 650 nights · 1 body.

The threshold falls as nights accumulate: more personal context, tighter detection. At 168 nights, the threshold reaches 18.53%. Half a year of data and you can detect smaller shifts that would have been indistinguishable from noise at 28 nights. The table does not promise the effect exists. It says when you could see it, if it did.

Why the null comes first

The house holds 18,483 measurements of one body. Sleep latency — median 11 minutes, coefficient of variation 84.41% — is among the noisiest metrics in the dataset. Any compound claiming to shorten sleep onset time is working against that noise floor before any other consideration applies.

Ben Maaoui's trial chose the most favorable possible population: people already sleeping well, measured with objective actigraphy. The compound produced no useful change in onset latency. That result is not a footnote. It is the central finding of this record. Published in full, with statistics, before anyone asks. A null with numbers is more useful than a claim without evidence behind it.

The limit

What this data does NOT say:

The null may be false by statistical power. n=14 is insufficient for detecting small effects. A non-significant result at that sample size does not prove the effect is absent. It proves the study lacked the statistical power to find it.

Loading dose is not maintenance dose. The protocol used 20 g per day for 7 days. Most users take 3–5 g per day on an ongoing basis. Effects at maintenance doses may differ from loading-phase results.

This population is not yours if you sleep poorly. PSQI below 5 means subjects were already good sleepers before the trial began. There is no data from this study on creatine and sleep latency in people with difficulty falling asleep or poor baseline sleep quality.

The devices are not interchangeable. The study used an ActiGraph GT3. The house dataset uses Oura. Different sensors, different analysis windows, different algorithms. Absolute values between devices are not directly comparable.

One body is not your body. The thresholds in this table — 45.38% at 28 nights, 18.53% at 168 nights — are calibrated to the house dataset: 650 nights of one subject measured on Oura. Your variability may be higher or lower. The instrument estimates your own thresholds from your data once you have accumulated enough personal context.

---

• Mental Work — all protocols
• related file: magnesium-for-sleep
• open science — the dataset

Dataset: 650 nights · 31 metrics · 18,483 measurements · one body · verified.