Analyses /Glucose by the clock

Glucose by the clock

2026-08-15 → 2026-08-28   exploratory — not registered in advance
“How does my glucose vary as a function of time of day?”
asked by Alan, before any analysis was run

3,923 continuous-glucose readings over 14 days in August 2026, asked what they do as a function of time of day. They are strongly clock-aligned: about 107 mg/dL between the 05:00 trough and the 11:00 peak. He falls through the night, climbs steeply after 08:00, and shows no early-morning rise at all. Every result here is a lead rather than a finding, and the last section says what would change that.

Results

Five sections: the shape of his day, three ways of measuring it, what was done to try to break the result, the overall level for context, and one thin lead reported with its fragility.

The average day

Is glucose aligned to the clock at all? Yes.

His glucose is lowest around 05:00 and highest around 11:00 — a swing of about 107 mg/dL across the day.

night 175 200 225 250 275 300 325 180 mg/dL peak 11:00 · 293 trough 05:00 · 194 00 03 06 09 12 15 18 21 days hour of day (local wall-clock) glucose (mg/dL)

Each point is the mean across the 14 days of that clock hour's readings. The band is a 95% bootstrap interval <b>resampling days</b> — how much this shape would move had he worn the sensor on a different fortnight. Dexcom's own report shades the spread of readings instead, which looks similar and answers a different question. 107 mg/dL against a null that reaches 63 only one time in twenty (p&lt;0.0001, 13 fully covered days).

hourglucoseinterval
trough05:00189 mg/dL175 – 214
peak11:00296 mg/dL271 – 315
swing107 mg/dLnull median 43, 95th pct 63

<b>The shape has two parts.</b> Overnight he <i>falls</i> — from 244 at midnight to 194 by 05:00, steadily, with no early-morning rise. Then from 08:00 he climbs hard, gaining 95 mg/dL in four hours to the 11:00 peak. A second, smaller rise arrives around 21:00.

<b>This shows <i>when</i>, not <i>why</i>.</b> The late-morning peak is almost certainly breakfast and the evening bump almost certainly dinner, but nothing in this export records a meal, so both are guesses. Logging meal times for two weeks would settle it, and is the single most valuable change available.

Three ways of asking the same question

How big is the day–night difference, and is there a dawn effect? Big, and no.

His waking day averages 49 mg/dL above his night. But there is no early-morning rise: between 03:00 and 08:00 he is still drifting <i>down</i>.

-0 25 50 75 no difference Waking day (08–22) − night (00–06) +49.1 n=13 days Dawn rise (06–08 mean − 00–05 trough) +7.4 n=14 days Dawn slope, 03:00–08:00 (per hour) -1.4 n=14 days effect (mg/dL, or mg/dL per hour)

Each row is one within-day contrast, tested across days with a sign-flip permutation. Intervals are normal approximations on the day-level standard deviation; the p-values are the exact permutation values. The day–night gap is large and reliable. The dawn slope is -1.4 mg/dL per hour with p=0.24 — reported as a null, because a tested absence is worth more than a silence.

contrasteffectdaysp
Waking day (08–22) − night (00–06)+49.1130.0010effect found
Dawn rise (06–08 mean − 00–05 trough)+7.4140.0015effect found
Dawn slope, 03:00–08:00 (per hour)-1.4140.2437no effect found

<b>The absent dawn effect is the interesting one.</b> A classic &ldquo;dawn phenomenon&rdquo; is a rise in the hours before waking, and it is what the first subject in this project shows. Alan does the opposite: he keeps falling until roughly 08:00, and his rise begins after that. Whatever drives his mornings starts when the day does, not before it.

What was attacked

Could the sensor itself have produced this? No.

The shape survives dropping either sensor, and survives discarding the first 24 hours of each one.

175 200 225 250 275 300 325 sensor 1 sensor 2 00 03 06 09 12 15 18 21 hour of day (local wall-clock) glucose (mg/dL)

The average day computed separately on each of the two transmitters. The transmitter changed once, mid-window, so these are two separate pieces of hardware measuring the same person — a stronger split than a calendar one, because it varies the instrument as well as the days. The swing ranges 97–116 mg/dL across the two leave-one-out fits and the peak does not move from 11:00.

attackresultp
drop sensor 1swing 97 mg/dL, peak 11:000.00625effect found
drop sensor 2swing 116 mg/dL, peak 11:000.00050effect found
drop 24 h after each sensor startday−night +47 mg/dL (12 days left)0.0042effect found
is it just his level, not a shape?r=-0.81 against his overnight mean

<b>Dexcom records the hardware change that Libre's export does not.</b> The transmitter id changes when the sensor does, so these wear periods are known rather than inferred from gaps — which is why this check is stronger here than the equivalent one on the first subject, where sensor boundaries had to be guessed. The day-night gap is larger on days he runs lower overnight (r=-0.81), so it is not a restatement of his level.

Two sensors is the smallest number that permits this test at all. It rules out one sensor having invented the pattern; it cannot rule out a fortnight that was unusual for other reasons.

What the fortnight looked like overall

The level, stated plainly and not interpreted

Average 242 mg/dL over the 14 days, with 12% of readings between 70 and 180.

night 175 200 225 250 275 300 325 14-day average 242 00 03 06 09 12 15 18 21 days hour of day (local wall-clock) glucose (mg/dL)

The same average day, with his own fortnight mean marked instead of the 180 line. Every figure reproduces Dexcom's own report to within 0.4 — which is what says the file is being read correctly.

figurethis analysisDexcom Clarity
mean mg/dL241.9242.0
gmi %9.19.1
sd mg/dL51.051.0
cv %21.121.1
in range 70 180 %12.412.0
low 54 69 %0.00.0
very low %0.00.0
very high %43.644.0
high 181 250 %43.944.0

<b>These numbers are reported, not interpreted.</b> This apparatus reports associations; what a glucose level means for a person is a clinical question and belongs to his doctor, not to this page. The figures are here because they are his, because they are the context for everything above, and because reproducing the vendor's own arithmetic is how a reader knows the parsing is right.

Weekends

Does the day of the week matter? Possibly — but do not act on it yet.

Weekend days averaged 29 mg/dL lower than weekdays (p=0.02) — on 4 weekend days.

200 250 250 mg/dL weekday n = 10 222 mg/dL weekend n = 4 whole-day mean (mg/dL)

One dot per day. Bars are 95% bootstrap intervals on each group's mean. Statistically this clears the usual threshold. Practically it rests on 4 weekend days inside a single fortnight, which is not enough to separate &ldquo;weekends are different&rdquo; from &ldquo;those two weekends were different&rdquo;.

daysmean glucose
weekday10250 mg/dL
weekend4222 mg/dL

<b>Reported because it was tested, flagged because it is thin.</b> This was one of several exploratory comparisons and it has not been corrected for that. The honest status is a lead worth re-testing on the next sensor, not a finding — and it is exactly the kind of result a longer wear would settle cheaply.

What this first pass deliberately does not do

Relevant data sources — timeline

When each stream of data actually exists. A stream that stops before the analysis window opens is greyed out: it is in the files, but it cannot speak to this question.

analysis window DEXCOM G7 CGM readings 3,923 readings SENSORS sensor 1 7 days sensor 2 8 days

One continuous fortnight of wear across two transmitters. The final day ends mid-afternoon, where the export was taken.

Relevant data sources — description

What each source contributes to this question — and, where it matters more, what it cannot contribute. Every item links back to the audit page for the file it comes from.

Dexcom G7 (via Dexcom Clarity)1 outcome · 1 covariate · 1 excluded

A Clarity CSV export covering 2026-08-15 to 2026-08-28. Names and device serials were removed before analysis; timestamps and glucose values are untouched. See the data audit →

Outcome

Glucose Value (mg/dL)

rows where Event Type is EGV · 3,923 readings, 5-minute cadence · sensor

Validated against Dexcom's own Overview report: mean 242 vs 242, GMI 9.1 vs 9.1, CV 21.1 vs 21.1.

Covariate

Transmitter ID

on every EGV row · 2 transmitters, changing 2026-08-21 · sensor

Pseudonymised to T1/T2 rather than deleted, which preserves the sensor boundary. That is what makes the leave-one-sensor-out check possible.

excluded

Alert / Device / Sensor rows

non-EGV rows · 18 rows · self-reported

His alarm thresholds and device models. Not measurements — a glucose column on an Alert row is a setting, and reading the file without the EGV filter would treat it as a reading.