How much does sleep latency have to change before it stops being noise? The threshold from 650 nights

To cross the noise floor of sleep latency, a change has to exceed 45.37% over a 28-night window. At 56 nights, the threshold drops to 32.08%. At 84 nights, 26.2%. At 168 nights — roughly half a year of continuous recording — it reaches 18.52%. These thresholds come from 650 nights of a single body, with a single device, without interruption. They are not theoretical estimates: they describe the measured behavior of one real variable.

The number your device skips

Sleep latency — the time between lying down and falling asleep — moves with enormous amplitude from one night to the next. The coefficient of variation (CV) of this metric in the dataset is 84.41%. CV measures the typical scatter of nightly values around the median as a fraction of the median itself. A CV of 84.41% means the usual dispersion is nearly as large as the central value.

The median across 650 nights is 11 minutes. Nights that sit far from that value are not outliers: they are the expected behavior of this variable.

The autocorrelation (rho) is 0.026 — effectively zero. Autocorrelation measures how well today's value predicts tomorrow's. At 0.026, the answer is almost not at all. Last night's latency carries almost no information about tonight's. Each reading is, statistically, nearly independent of the one before it.

The data

This table comes directly from ruido.medido.json#sleep_latency_min. No values have been adjusted or rounded.

Parameter	Value
Nights in the record	650
Median	11 min
CV (relative variability)	84.41%
Rho (autocorrelation)	0.026
RCV (night-to-night threshold)	233.96%
Threshold at 28 nights	45.37%
Threshold at 56 nights	32.08%
Threshold at 84 nights	26.2%
Threshold at 168 nights	18.52%

What the RCV of 233.96% means

The RCV — Reference Change Value — is the threshold between two individual measurements. It answers how different tonight's latency has to be from last night's to be distinguishable from the variable's normal fluctuation.

For sleep latency, that threshold is 233.96%. In practical terms: comparing last night's latency to tonight's is, statistically, noise reading noise. This is not a flaw in the device. It is the honest description of what this variable does.

The device shows you the number. What it does not show is whether that number has actually moved — or whether it is still within the range it always occupied, before you changed anything.

How the thresholds shift with accumulated nights

The alternative to comparing individual nights is to accumulate. As the measurement window grows, the standard error of the estimate decreases and the detectable threshold shrinks.

At 28 nights, the threshold is 45.37%: a change of that size in average latency between one 28-night period and the next would be detectable as a real signal. At 56 nights: 32.08%. At 84 nights: 26.2%. At 168 nights — roughly half a year of continuous recording — the threshold reaches 18.52%.

Accumulation does not change the underlying biology or the intrinsic variability of the variable. It changes the resolution of the instrument. More nights in the window yields a more precise estimate of the true level of the metric, which makes smaller changes detectable.

The pattern is straightforward: sleep latency is a variable that demands time to become readable. At 28 nights, the instrument can see something. At 168 nights, the threshold has dropped substantially. The instrument sharpens — but sharpening costs nights.

Where the 650 nights come from

The dataset holds 650 nights of a single body, 31 metrics, and 18483 measurements. Sleep latency is one of those 31 variables in continuous record. Other metrics in the same dataset have substantially lower CVs than 84.41%, which translates to smaller detection thresholds and readable signals from shorter windows.

Sleep latency sits at the volatile end of the spectrum. Knowing that position does not invalidate the metric: it frames it. Knowing that the 28-night threshold is 45.37% is information the wearable does not surface. The wearable records the number; the record builds the framework to read it.

The 650 nights are from one body, one device, one continuous period. There is no second subject in this record for this metric. There is no comparison group. What exists is enough history of a single variable to describe its own behavior accurately.

The limit

This data answers how much sleep latency has to change to exit the noise floor. It does not answer why it changes, or what would drive it to change in a given direction.

The thresholds — 45.37% at 28 nights, 32.08% at 56 nights, 26.2% at 84 nights, 18.52% at 168 nights — are values computed from one body's record during a specific period. They may differ in another body, with a different device, or at a different point in the same person's life.

The CV of 84.41% combines real biological variability with sensor noise. Separating the two would require simultaneous measurement with two different devices on the same nights — a condition this dataset does not meet for sleep latency.

Sleep latency is not the same construct across all wearables. The sleep-onset detection algorithm, the analysis window, and the internal thresholds differ between manufacturers. The figures in this row belong to the case zero record with its specific device. Applying them to a different device without building the same record from the beginning would be extrapolation without data support.

A change that falls short of 45.37% over a 28-night window is not necessarily absent biologically. It may be real and below the instrument's resolution at that number of accumulated nights. The threshold describes when the instrument can see — not when something is happening.

---

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