How much does bedtime start min have to change before it stops being noise? The threshold from 650 nights

Across 650 nights from a single body, a change in bedtime_start_min has to pass 18.43% between two 28-night windows before it stops being noise, or 13.03% when comparing two 56-night windows. Below those thresholds, the variation you are looking at fits inside what this variable already moves on its own, night after night, without anything actually happening.

What is bedtime_start_min, and why does it need its own threshold?

bedtime_start_min is the minute of the day when recorded sleep begins. Like every sleep metric, it is not stable: it drifts for reasons that rarely have a name — a night with traffic, a show that ran long, two days of jet lag. That background drift is what this house calls noise, and every variable carries its own amount of it. Asking "did bedtime_start_min go up this week?" without knowing its background noise is a question with no possible answer: whatever number comes out could be signal, or it could be the same noise as always, dressed up as a trend.

The coefficient of variation (CV) of bedtime_start_min in this body is 33.84%. That is a high number next to more stable metrics from the same body, and it says something simple: night to night, bedtime_start_min moves a lot on its own, with no intervention from anyone. That 33.84% is the starting point for calculating how much a multi-night average — not a single night — has to move before the movement stops being explainable by baseline noise.

How it was measured: 650 nights, one body, no population averages

The number comes from 650 consecutive nights of a single body, recorded inside this house's protocol. It is not an average across thousands of users of a sleep app, nor a figure from a paper with a large n and conditions that never quite match yours. It is the real variability of one person, night after night, across 650 nights — part of a corpus of 18483 measurements spanning 31 metrics that backs the entire protocol.

From those 650 nights, the raw CV (33.84%) was calculated first, and then the thresholds that actually matter for deciding anything: how far apart the average of one window of nights has to sit from another window before the difference is, with high probability, not just the same random variation as always. These thresholds are calculated for four window sizes — 28, 56, 84, and 168 nights — because averaging over more nights reduces noise, which is why the threshold drops as the window grows.

The table: the threshold by how many nights you average

Window (nights)	Real-change threshold
28	18.43%
56	13.03%
84	10.64%
168	7.52%

Reading the table is direct: if you compare the average bedtime_start_min of the last 28 nights against the 28 nights before that, you need a change larger than 18.43% for it to be worth looking at. If instead you have the patience to compare 168-night windows, the threshold drops to 7.52% — because averaging 168 nights dilutes much of that 33.84% swing that dominates on a night-by-night basis.

This pattern — the threshold dropping as the window grows — is not a quirk unique to bedtime_start_min. It is arithmetic of averages: the more data sits on each side of the comparison, the less weight a single odd night carries. What is specific to bedtime_start_min is the starting point, that 33.84%, along with the fact that this variable moves largely out of step with other metrics from the same body: the measured correlation coefficient (rho) is 0.039, close to zero. The relative cross-correlation (rcv) is 93.81%, another number confirming the same thing from a different angle — bedtime_start_min mostly does its own thing.

Why this matters before drawing a conclusion

The temptation with any sleep metric is to look at last night's number, compare it to the night before, and draw a conclusion about what moved it. With a CV of 33.84%, that single-night comparison sits, almost always, inside the noise — it says nothing about whether anything real changed. The 28-night threshold (18.43%) is already a huge jump in reliability compared to comparing two lone nights, but it remains the most demanding of the four thresholds precisely because its window is the shortest.

Anyone who wants to detect a real change in bedtime_start_min with the shortest possible wait should work with windows of at least 28 nights and require the change to pass 18.43%. Anyone with room to wait longer — and who wants a less demanding threshold because they are averaging over more nights — can move to 56, 84, or 168 nights, where the threshold drops as low as 7.52%.

The limit

This data does not say what caused a change in bedtime_start_min, nor whether that change is good or bad for anyone. It is a statistical threshold calculated over 650 nights from a single body — mine — inside one specific protocol. It has not been validated across other bodies, other measurement devices, or sleep routines different from the one recorded here. The rho of 0.039 shows that bedtime_start_min behaves almost independently of other variables in the same protocol, but that does not mean it is independent of outside factors this corpus never measured — ambient noise, seasonal shifts, travel. Nor does it say whether a change above the threshold is desirable: it only says that change is large enough to not be explained by this metric's own background variation, in this body. And the threshold itself is specific to bedtime_start_min: the same number cannot be applied to another variable in the protocol without repeating this calculation on its own row of data.

Keep looking

This threshold is one piece inside a wider protocol for measuring the body during mental work and rest. You can browse Mental Work — all protocols to see the rest of the measured variables, the related file: magnesium-for-sleep for a sleep-specific protocol, or go straight to open science — the dataset to see the 650 nights and 18483 measurements behind this number.