Skip to content
The science

Cortisol reference curves: how the day-curve is built.

The curve on our cortisol meter page is built from other people's published research data, not from the Auromone Curve. This page lists every study behind it, how the numbers were put on one scale, what was left out, and where the result should not be trusted.

The short answer

The Auromone cortisol reference curve is a population backdrop, not a healthy range. It plots 76,281 readings from 3,710 adults in 10 published studies, harmonized to a salivary-equivalent scale in nmol/L, as a median with the middle 50% and 80% of readings for each half hour of the clock. Only aggregate statistics are published. It cannot tell anyone whether a result is normal.

What is the cortisol reference curve?

It answers one question: across a large number of adults sampled in daily life, where did cortisol readings fall at each time of day? It feeds two things on this site, the interactive day-curve you can filter by sex and age, and the free cortisol levels chart that writers can reuse.

Readings76,281
Participants3,710 adults
Source studies10
Strata18: all adults, men and women, each at all ages and in five age bands
Time resolutionHalf-hour bins, 00:00 to 24:00 (49 bins)
Statistics per bin10th, 25th, 50th, 75th and 90th percentiles, plus the number of readings
UnitSalivary-equivalent cortisol, nmol/L

The all-adults curve at selected times

Clock time10th25thMedian75th90thReadings
01:000.40.61.01.62.8207
06:307.011.217.426.134.73,090
08:007.210.515.823.733.04,461
12:002.84.57.412.218.54,776
16:001.52.34.06.911.4440
21:000.81.32.95.49.24,489
23:000.71.32.85.49.02,059

Percentiles in nmol/L, after display smoothing. Readings are the count in that half-hour bin. A population reference is a backdrop, not a target. These are not diagnostic thresholds.

Which studies feed the curve?

Ten sources, each listed with who took part, how they were sampled, how cortisol was assayed, and how many people and readings reached the final curve. These counts are what survived our filters, so they can differ from the headline number in a paper.

SourceWhoSamplingAssayPeopleReadings
MIDUS 2
Data · Paper
U.S. national sample, Daily Stress Project, 2004 to 2009. Ages 33 to 84Four saliva samples a day on 4 days: at waking, 30 minutes after getting out of bed, before lunch, before bedLuminescence immunoassay1,73025,465
MIDUS 3
Data
Returning MIDUS 2 diary participants, 2017 to 2019. Ages 43 to 90Same protocol as MIDUS 2Luminescence immunoassay1,00514,453
MIDUS Refresher
Data
New U.S. sample, 2012 to 2014. Ages 25 to 75Same protocol as MIDUS 2Luminescence immunoassay6689,507
ULTRADIAN
Paper · Data
Healthy volunteers at sites in Bristol, Bergen, Stockholm and Athens. Ages 20 to 69Free cortisol in tissue fluid by microdialysis over 24 hours, in 20-minute samples, including sleepLC-MS/MS21314,732
Well Elderly 2
Trial · Cortisol paper · Data
Independently living older adults in Los Angeles, 2004 to 2008. Ages 60 to 90 in our file. A lifestyle randomized trial: see the inclusion sectionFour saliva samples per assessment: at awakening, 30 to 45 minutes later, 5 hours later, 5 hours laterEnzyme immunoassay4054,554
heiDATA
Paper · Data
Community adults in Germany during the first COVID-19 lockdown, April to June 2020. Ages 18 to 70Six saliva samples a day on 2 days: at awakening, 30 minutes, 45 minutes, 2.5 hours and 8 hours after, and before sleepELISA2452,791
Bierstetel
Paper · Data
Married couples. The paper analysed 82 people; the deposited file holds more18 saliva samples over 3 daysNot confirmed*971,513
Booij (MOOVD)
Paper · Data
Non-depressed control participants only, ages 20 to 50. The depressed arm is excludedThree saliva samples a day for 30 days, about 6 hours apartLC-MS/MS151,321
CARWatch
Paper · Data
Healthy participants, mean age 24Saliva samples after waking on 2 consecutive daysNot confirmed*1171,122
Haase, Study 2
Paper · Data
Healthy naturally cycling women. Ages 18 to 40Saliva samples after waking on 7 consecutive days around expected ovulationNot confirmed*30823

People add up to 4,525 across rows and 3,710 unique, because 815 people took part in both MIDUS 2 and MIDUS 3. *We could read only the abstract of these three papers, which does not state the assay, so we do not claim one. Together they supply 3,458 readings, 4.5% of the total. Of the rest, 56,770 readings (74.4%) came from immunoassays and 16,053 (21.0%) from LC-MS/MS.

How were the values harmonized?

Saliva values were left as reported. One source was rescaled, and it is the only rescaling on the page. The steps below run in this order.

  1. Units. Everything is in nmol/L. Most sources report nmol/L directly. Bierstetel's deposited file holds natural-log values, which we back-transform. For CARWatch, Haase and Bierstetel the unit was confirmed from the range of the values rather than from a codebook.
  2. One scale. ULTRADIAN measured free cortisol in tissue fluid, not saliva. Two constants bring it across: ×2.47 for incomplete microdialysis recovery, then ×0.45 from free cortisol to saliva, a net ×1.11. Both are our own fitted estimates, not published conversion factors. The 2.47 was set from where ULTRADIAN overlaps the saliva curves between 15:00 and 20:00, so agreement in that window is by construction and is not independent validation.
  3. Clock time. A recorded clock time is used wherever one exists: all of ULTRADIAN and 91% of MIDUS, or 59,909 readings (78.5%). The other 16,372 sit on a nominal 07:00 wake plus their hours since waking. That covers Well Elderly 2, heiDATA, Bierstetel, Booij, CARWatch, Haase and the rest of MIDUS.
  4. Named occasions. Samples recorded only as an occasion take a fixed offset from waking: 5 hours for midday or before lunch, 14 hours for bedtime. Well Elderly 2's evening sample sits at 10.5 hours, which matches that study's recorded median of 17:30. Booij's three day-parts sit at 2.5, 8 and 14 hours.
  5. Binning. Each reading goes to the nearest half hour.
  6. Percentiles. Nearest-rank percentiles per bin. Readings are not weighted by study or by person.
  7. Thin bins. Any bin with fewer than 10 readings in a stratum is dropped from that stratum.
  8. Smoothing. Each percentile gets a light 0.25, 0.5, 0.25 weighting across neighbouring bins. This changes how the line is drawn, not any measurement.

We tested rescaling each study to a common level and rejected it. It made the line tidier by altering real measurements, so differences between assays stay in the data and appear below as a limit instead.

What did quality control remove?

StepRuleReadings left
IngestedAll rows from the sources above, plus the Booij depressed arm83,173
StructuralDrop empty, non-numeric, zero or negative values81,421
Plausible boundsKeep saliva from 0.1 to 50 nmol/L and tissue fluid from 0.05 to 100 nmol/L79,756
StatisticalTukey fences at 1.5 times the interquartile range on log values, within each source and time bin, for groups of 20 or more77,651
Clinical splitMove the Booij depressed arm out of the reference (1,312 readings)76,339
Usable timeDrop readings that cannot be placed on the clock76,281

Raw MIDUS values reached about 3,000 nmol/L, which is why hard bounds exist. They cut both ways: a genuine very high reading is removed along with the errors, so the top of the distribution is understated.

Who is included, and who is left out?

A dataset qualifies if it meets these rules, checked in this order:

  • Not patients. Adults from general or community samples, not recruited for a disease.
  • Daily life. Cortisol sampled naturalistically. No stress task, drug challenge or other provoked response.
  • Known timing. A clock time, a time since waking, or a documented schedule.
  • Known units. Ambiguous units mean the dataset stays out.

"Healthy" here means "not recruited as a patient", and that is a weaker promise than it sounds. MIDUS and Well Elderly 2 sampled the community, and we apply no health or medication filter inside those cohorts. Published MIDUS analyses include participants who used medications known to affect cortisol and who reported poorer health. Booij, ULTRADIAN, CARWatch and Haase recruited healthy people specifically.

Two sources are exceptions we kept on purpose. Well Elderly 2 is a lifestyle randomized trial, and our file includes assessments taken after the intervention. CARWatch is a methods study of an app for timing saliva samples. Both sampled naturalistically, and their normalized curves stayed close to the pooled shape, with a mean difference of 5.5% and 11.7%, so they stayed in. A stricter reader could drop them.

Left out, with the reason:

  • Participants with depression in the Booij study, held separately from the reference.
  • Stress-task and challenge studies, because they measure a provoked response rather than a daily rhythm.
  • Children's data, because the curve describes adults.
  • Datasets with reconstructed sample times or ambiguous units, because either would corrupt the axes.
  • A wake-aligned release of ULTRADIAN, because it would count the same people twice.

How many people are in each stratum?

The curve can be filtered by sex and age, which gives 18 strata. A race and ethnicity filter was removed in August 2026 because coverage was too uneven to compare groups fairly.

Age bandAllMenWomen
All ages3,7101,4792,133
Under 35526189337
35 to 49934384549
50 to 641,384610774
65 to 74854363491
75 and over476181295

Participants per stratum. Age is taken at each reading, so a person sampled in two waves can appear in two bands and the bands add up to more than the total. Readings with no recorded sex (1,533) or age (1,539) count only toward "All". Nearly all of them are Bierstetel, whose file gave us neither.

What are the limits of the curve?

  • It is not a reference interval. It has not been peer reviewed or validated for any clinical use. Being inside a band does not mean a result is normal, and being outside one does not name a condition. Doctors confirm or exclude a diagnosis with dedicated protocols, as set out in why cortisol disorders are hard to diagnose.
  • The upper bands run far above lab ranges. At 7 to 8 a.m. the 90th percentile is about 33 nmol/L, roughly 1,200 ng/dL, against the 750 ng/dL upper limit from University of Michigan MLabs. At 11 p.m. it is 9.0 nmol/L against under 100 ng/dL, about 2.8 nmol/L. Assay differences and the make-up of these samples may both contribute, and we cannot separate them. If a lab flags your result, that flag stands.
  • The people behind the line change by the hour. Between 00:30 and 03:30, ULTRADIAN supplies 69% to 100% of readings, so the overnight floor is largely one method in about 200 people. It is also the only source that sampled during sleep.
  • Clock time blurs the morning. The rise after waking follows each person's wake time, not the clock, so the curve shows a broad hump rather than the sharp cortisol awakening response one person has.
  • Some placement is approximate. About a fifth of readings sit at a fixed offset from a nominal 07:00 wake rather than at the time that person sampled. Booij's samples sit at nominal day-parts, and every Well Elderly 2 evening sample sits at one study-wide time, the 17:30 median, although most recorded times run from about 16:30 to 19:30.
  • It pools readings, not people. ULTRADIAN gives about 69 readings per person and Booij about 88, against about 15 in MIDUS 2. The bands therefore mix differences between people with one person's movement across days.
  • It reflects who was studied. Most readings are from U.S. adults. Where race was recorded, about 85% of readings came from people recorded as White. heiDATA was collected during a lockdown. No one under 18 is included.
  • It does not separate groups that matter. Shift work, pregnancy, hormonal contraception, adrenal conditions and steroids taken by any route are not modelled, so the curve should not be read against any of them.
  • It is a saliva scale. It is not blood, and it is not sweat. Blood reports mostly protein-bound total cortisol, so the two cannot be compared side by side. The best time to test cortisol depends on the test being used.

Why are only aggregates published?

No individual-level row leaves our pipeline, and that will not change. Source terms commonly forbid redistributing participant records, so only aggregates ship. The published aggregate file holds, for each stratum, a count of people and readings and, for each half-hour bin, five percentiles and a reading count. It holds no participant IDs and no bin with fewer than 10 readings. A reproducible script builds it from the cleaned dataset, and rerunning that script on 6 November 2026 gave a byte-identical file.

The curve is also not Curve data. Nothing on this page describes how the Auromone Curve performs, and we publish no accuracy figure for it.

Data credits and how to cite

The data belong to the studies that collected them. The original collectors, ICPSR and the funding agencies bear no responsibility for this use of the data or for the interpretations on this page.

  • Ryff, Carol D., and Almeida, David M. Midlife in the United States (MIDUS 2): Daily Stress Project, 2004-2009. Inter-university Consortium for Political and Social Research [distributor], 2017-11-20. doi.org/10.3886/ICPSR26841.v2
  • Ryff, Carol D., and Almeida, David M. Midlife in the United States (MIDUS 3): Daily Diary Project, 2017-2019. Inter-university Consortium for Political and Social Research [distributor], 2022-09-22. doi.org/10.3886/ICPSR38529.v1
  • Ryff, Carol D., and Almeida, David M. Midlife in the United States (MIDUS Refresher 1): Daily Diary Project, 2012-2014. Inter-university Consortium for Political and Social Research [distributor], 2020-12-14. doi.org/10.3886/ICPSR37083.v2
  • Clark, Florence. Well Elderly 2, Los Angeles, California, 2004-2008. Inter-university Consortium for Political and Social Research [distributor], 2013-05-28. doi.org/10.3886/ICPSR33641.v1
  • Almeida DM, Piazza JR, Stawski RS. Interindividual differences and intraindividual variability in the cortisol awakening response. Psychology and Aging. 2009;24(4):819-27. doi.org/10.1037/a0017910
  • Dmitrieva NO, Almeida DM, Dmitrieva J, Loken E, Pieper CF. A day-centered approach to modeling cortisol. Psychoneuroendocrinology. 2013;38(10):2354-65. doi.org/10.1016/j.psyneuen.2013.05.003
  • Upton TJ, Zavala E, Methlie P, et al. High-resolution daily profiles of tissue adrenal steroids by portable automated collection. Science Translational Medicine. 2023;15(701):eadg8464. doi.org/10.1126/scitranslmed.adg8464. Replication data on DataverseNO, CC0: doi.org/10.18710/5TW8YF
  • Clark F, Jackson J, Carlson M, et al. Effectiveness of a lifestyle intervention in promoting the well-being of independently living older people: results of the Well Elderly 2 Randomised Controlled Trial. J Epidemiol Community Health. 2012;66(9):782-90. doi.org/10.1136/jech.2009.099754
  • Wilcox RR, Granger DA, Szanton S, Clark F. Diurnal patterns and associations among salivary cortisol, DHEA and alpha-amylase in older adults. Physiology & Behavior. 2014;129:11-6. doi.org/10.1016/j.physbeh.2014.02.012
  • Hopf D, Schneider E, Aguilar-Raab C, et al. Loneliness and diurnal cortisol levels during COVID-19 lockdown. Scientific Reports. 2022;12:15076. doi.org/10.1038/s41598-022-19224-2. Data on heiDATA: doi.org/10.11588/DATA/SYVQMM
  • Bierstetel SJ, Slatcher RB. Couples' behavior during conflict in the lab and diurnal cortisol patterns in daily life. Psychoneuroendocrinology. 2020;115:104633. doi.org/10.1016/j.psyneuen.2020.104633. Data on OSF: osf.io/ynwmr
  • Booij SH, Bos EH, Bouwmans MEJ, et al. Cortisol and α-amylase secretion patterns between and within depressed and non-depressed individuals. PLOS ONE. 2015;10(7):e0131002. doi.org/10.1371/journal.pone.0131002. Data on DataverseNL: doi.org/10.34894/TCFHXT
  • Richer R, Abel L, Küderle A, Eskofier BM, Rohleder N. CARWatch: a smartphone application for improving the accuracy of cortisol awakening response sampling. Psychoneuroendocrinology. 2023;151:106073. doi.org/10.1016/j.psyneuen.2023.106073. Data on OSF: osf.io/c35f2
  • Haase L, Vehlen A, Strojny J, Domes G. Effects of menstrual cycle phase and ovulation on the salivary cortisol awakening response. Psychoneuroendocrinology. 2024;160:106669. doi.org/10.1016/j.psyneuen.2023.106669. Data on OSF: osf.io/eyhuj

To cite this page: Auromone. Cortisol reference curves: method, sources and limits. Updated 6 November 2026. auromone.com/science/cortisol-reference-curves. The cortisol levels chart is free to reuse under CC BY 4.0, with embed code on that page.

This page is for general education and is not medical advice. The Auromone Curve is designed as a general wellness device, not a diagnostic, and does not replace clinical testing. If you are concerned about your cortisol levels or a test result, talk to a healthcare provider.

Straight answers

Reference curve FAQ

Is the cortisol reference curve a normal range?

No. It shows where other adults' readings fell at each time of day, pooled across studies with different assays and community samples that include people with health conditions and on medication. Being inside a band does not mean a result is normal, and being outside one does not name a condition. Only a clinician reading a result against the range on that lab's report can judge it.

Does the reference curve come from the Auromone Curve?

No. Every reading comes from published research studies that sampled saliva or tissue fluid. None comes from the Auromone Curve, which is a general wellness device in development and designed to read cortisol from sweat. The reference says nothing about how that device performs.

Why do the curve's numbers differ from lab reference ranges?

Because they measure different things in different ways. Most readings behind the curve came from immunoassays in general community samples, while a lab range is tied to one validated method and a defined reference group. The upper band of the curve sits well above the saliva range published by University of Michigan MLabs. Use the range on your own report, never the curve.

Why does the curve peak between 6 and 7 a.m. rather than at waking?

Because it is drawn by clock time, and people wake at different times. Studies with recorded clock times spread their morning peak across the early hours, and studies that recorded only time since waking are placed on a nominal 7 a.m. wake. That blurs the rise after waking into a broader, lower hump than any one person shows.

Why is there no filter by race or ethnicity?

Coverage was too uneven to present as a comparable filter. Race was recorded for about two thirds of readings, and about 85% of those came from people recorded as White, so several groups would have rested on very few people. Auromone removed the filter from the public curve in August 2026.

Can I reuse the cortisol chart?

Yes. The cortisol levels chart is free to republish under CC BY 4.0 with credit and a link to Auromone. That licence covers the chart. The underlying study data stay under each study's own terms.

Every line here is someone else's day.

These 76,281 readings came from other people's saliva and tissue fluid. The Auromone Curve is designed to read cortisol from a trace of sweat on your wrist, so the next curve you study can be your own. It is a general wellness device in development, it will not tell you whether a reading is normal, and it isn't on sale yet. Reserving is free.

Reserve