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

For a change in sleep efficiency to stop being noise and start being signal, it has to move by at least 2.43% across a 28-night window, or 1.72% across a 56-night window. That threshold comes from 650 nights of a single body, the same row that anchors the rest of this file. Below those figures, what you see in the morning app is not new information — it's the variation that metric already has when, in fact, nothing changed.

What sleep efficiency is, in the terms that matter here

Sleep efficiency is the share of time in bed that is actually spent asleep. It is one of the most cited metrics on any wearable, and also one of the most misread: a one-point jump from one day to the next gets read as an event almost every time, when it can simply be the normal range that variable moves in on its own. The question this file answers is not "did my sleep efficiency go up or down," but "did it go up or down by more than it moves on its own, for no apparent reason?"

That distinction can only be made with a body's own data and enough history behind it. There is no universal sleep efficiency threshold, because each person's background noise is different: it depends on the sensor, on routine, and on how much that specific person varies night to night. No one else's threshold gets imported here — it gets measured.

How it was measured: 650 nights of a single body

The file records 650 nights of sleep efficiency from a single body, inside a larger archive of 712 nights and 18483 total measurements across 31 variables. Across those 650 nights, the median sleep efficiency is 93%: half the nights in the file sit above that figure and half below it. That is the center of the distribution — the reference point against which any movement gets measured.

The coefficient of variation (CV) for the full series is 4.38%. That number summarizes how much sleep efficiency moves from one night to the next, relative to its own center, across the 650 nights. It's the base unit for everything else: if CV were zero, any change would be signal; at 4.38%, there's a band of normal movement that has to be subtracted before calling anything a real change.

The night-to-night correlation (rho) is 0.058, a number very close to zero. In practice: knowing last night's sleep efficiency tells you almost nothing about tonight's. Each night behaves, largely, like an independent event — which is exactly why a single bad or good night isn't enough to draw a conclusion, and why the threshold has to be calculated over windows of several nights, not a single isolated data point.

The table: the threshold changes with the window

The change threshold isn't a fixed number — it drops as the observation window grows, because averaging more nights reduces the noise carried by any single night. This table comes directly from the source row ruido.medido.json#sleep_efficiency, without rounding or reinterpretation:

Window (nights)	Minimum change threshold
28	2.43%
56	1.72%
84	1.4%
168	0.99%

Read another way: to know whether your sleep efficiency actually changed, comparing two 28-night blocks, that change has to clear 2.43% before you can rule out noise. Comparing two 168-night blocks — six months of history behind each — the threshold drops to 0.99%, because there's far more averaging absorbing the night-to-night variation.

Alongside the window thresholds, the source row also carries a relative noise measure (rcv) of 12.13%. It's another way of quantifying how much this variable moves relative to its own spread, and it confirms the same point as the rest of the table: sleep efficiency has real, measurable background noise, and it is not small next to the swings that usually get reported as day-to-day "improvements" or "declines."

Why this matters before reacting to a number

A threshold of 2.43% at 28 nights, or 1.72% at 56 nights, is not a statistical shrug — it's the line separating a morning reading from an actual trend. Most decisions people make about their sleep — switching a supplement, shifting a schedule, changing a routine — get triggered by a single number going up or down, without first asking whether that movement sits inside the range the variable produces on its own. With 650 nights of a single body, that range stops being a guess and becomes a number: a 4.38% CV, a 0.058 night-to-night correlation, and a table of window-based thresholds that says, precisely, how much it takes for a change to stop being noise.

The limit

This threshold is measured on a single body, from 650 nights of one data source. It does not say what caused any specific change in that person's sleep efficiency, and it does not establish a cause-and-effect relationship between any intervention and the metric. It also doesn't generalize to another body: another person's sleep efficiency background noise could carry a different CV, a different night-to-night correlation, and therefore a different threshold — even using the same sensor and the same definition of the metric. The threshold presented here is for reading this file with rigor; it does not replace measuring your own series if you want to apply the same standard to your own nights.

To continue

This file is one piece inside a larger archive. You can review the rest of the sleep thresholds at Mental Work — all protocols, the related file magnesium-for-sleep, or the full dataset at open science — the dataset.