Andy Garcia · freelance dev
FRENES
Back to the blog
Science & Data

300 Million Lab Tests Point to a Cliff at Menopause. Averaging by Age Had Turned It Into a Slope.

October 19, 2026 7 min read

On Monday, October 12, Nature Medicine published a paper from Uri Alon's lab at the Weizmann Institute of Science: "Population-scale reconstruction of menopause dynamics reveals a physiological cliff at the final menstrual period." The authors analyzed two datasets, each with its own version of the method. NHANES, the US health survey, from 1999 to 2022. And records from Clalit, the Israeli health insurer that covers about half the population, with 1.3 million women aged 30 to 70 between 2004 and 2024. Together, more than 300 million lab tests.

Plotted against age, most of those tests drift gradually through the forties and fifties. Plotted against time from each woman's final menstrual period, many of them jump. The paper reports step-like changes at that point in endocrine, bone, liver, kidney and lipid tests; no such jump in men run through the same analysis; jump sizes that agree between the US and Israeli data (r = 0.89); smaller jumps in women who reported hormone therapy; and hormone changes that start more than a decade before the final period.

I am not a clinician, and nothing here is health advice. I read the paper as someone who builds software and spends a lot of time staring at averages. The finding belongs to medicine. The method travels, and so do its limits.

How an average draws a slope

The problem is simple once you see it. Menopause does not happen at one age. In NHANES, the age at the final period has a median of 50 and an interquartile range of 45 to 53. So at 45, roughly a quarter of women are past it. At 50, half. At 53, three quarters.

Now imagine every woman had the same sudden step at her own final period. The average by age would only show the share of women who have crossed it: about a quarter of the step at 45, half at 50, three quarters at 53. A cliff, averaged across people who reach it at different ages, comes out as a ramp. The paper opens on exactly that point: menopause has mostly been described as a gradual transition, and using age instead of menopause timing "blurs the timeline."

The paper's example is ferritin, an iron marker. By age, it changes gradually. Realigned on the final period, it jumps by 0.6 standard deviations. Across the 90 tests in the US data, 52 show a significant jump in women compared with men, with sizes from 0.1 to 1.4 standard deviations. Most move in the direction of higher disease risk. The main exceptions are iron, hemoglobin and red blood cell tests, which improve, probably because menstrual bleeding stops.

What the method actually does

The authors call it deconvolution and cite a 1972 image-restoration paper: the way an astronomical image is sharpened by correcting for the distortion of the telescope. Here the blur is the spread of menopause ages.

The algorithm, which they named YoGlen, does not predict each woman's menopause date. It reconstructs the average trajectory as if everyone had been aligned on her own final period. In NHANES, menopause status is known for 73.2 percent of women, and 89 percent of those who said they were postmenopausal also gave an age. For the rest, the model draws 15 plausible menopause ages from the known distribution, consistent with what each woman reported, runs the analysis on each version and pools the results. Clalit has no menstrual data at all. There, each test is fitted with a simple model that includes a possible jump, averaged over a published distribution of menopause ages from 306,068 women.

The step is not forced. In the US analysis, each of 100 cross-validation repeats picks among models with and without a jump. In the Israeli one, the jump is a parameter estimated from the data.

Three checks make the result worth taking seriously. On simulated data, with the menopause distribution shifted by 4 years or made noisier, the method recovered the reference curves with correlations of 0.94 and 0.96. A direct comparison with no deconvolution at all, premenopausal versus postmenopausal women aged 45 to 53 with known status, matched the deconvolved jump sizes at r = 0.97. And men, pushed through the same machinery with the same distribution, show gradual change and no jump. That last one is the control I would copy first.

What it does not show

The headline number hides an asymmetry. The 300 million tests come mostly from Clalit, with a median of 3 million tests per lab marker. NHANES has a median of 14,778 people per test. The analyses that need menopause status, and the main hormone therapy comparison, run on the smaller survey.

The main analysis treats both datasets cross-sectionally and retrospectively, with self-reported menopause status in NHANES. It does not follow the same women through the transition; only the Clalit hormone therapy check compares the same women before, during and after treatment. The authors write that deconvolution "cannot replace the granularity of a prospective longitudinal study."

How sudden is sudden also stays open. In the paper's main figure, a jump size is a 3-year change, women minus men. Alon told Newsweek: "We can't say for sure how rapid it is but it looks like it's within the year of the final menstrual period."

The hormone therapy result is the one most likely to be over-read. Only 9 percent of postmenopausal NHANES participants reported it, and the matched analysis had 738 users. Thirty-nine of 80 tests differed. The authors adjusted for age, income, smoking, insurance, diabetes, education, ethnicity and survey period, and still write that this "cannot entirely eliminate healthy user bias," and that the effects "need to be tested by prospective randomized trials." Women on therapy also had higher CRP, an inflammation marker. All of it is association, not a recommendation.

The "decade earlier" finding comes from cross-sectional means from a few thousand women who said they had not reached menopause. AMH starts an accelerated fall about 15 years before the final period, FSH an accelerated rise about 14 years before, estradiol an accelerated drop about 12 years before. The Clalit data put the FSH and estradiol turns at nearly the same times. The authors place those two changes about 7 years earlier than previous studies.

Outside researchers did not all read it the same way. JoAnn Manson, a professor of medicine at Harvard Medical School, told Newsweek the study "confirms the profound physiological changes" women go through at the final period. Rebecca Thurston, of the University of Pittsburgh, told Newsweek that because the analysis is cross-sectional, the authors "cannot rigorously distinguish menopause (e.g., reproductive aging) from chronological aging," and that they "have very large amounts of missing data on key variables." And Martha Hickey, who heads menopause services at The Women's Hospital Victoria in Australia and was not involved, called the claim of "cliff-like" changes "unnecessarily alarming" for women, as quoted by the Pakistani newspaper Jang: "Physiological cliff is not a scientific or medical term."

And it is not brand new. A preprint has been on arXiv since November 8, 2025. The journal received the paper on January 11 and accepted it on July 28. The code is on GitHub, in R, with simulated data. The Clalit records are not public.

The same blur in my dashboards

Swap "menopause" for any event that happens to each person at a different time, and the trap is the one I watch for in product data. Subscriptions renew on each user's own date. Free trials end on each user's own day. In BidForge, the tool I built to read French public tender files for small businesses, every tender has its own submission deadline. Average activity by calendar week, and any rush before a deadline would smear into a gentle hill.

The opt-in problem is the same as the hormone therapy one. Users who switch on a feature often look healthier than those who do not, and they were often different before they switched. The paper's authors name it healthy user bias and ask for randomized trials. In software, the randomized trial is cheap: an A/B test.

What I would check

Before redesigning anything around a slow decline in a chart, I would:

  1. Re-plot the metric on time since the event, not on the calendar: days since signup, since trial end, days before a deadline.
  2. Run the same alignment on a group that has no event, the way the paper uses men. A step that also appears there is an artifact.
  3. Where the event date is missing, borrow the distribution from the users who have it, and validate on that subset, like the 45 to 53 check.
  4. Treat every "users who did X do better" chart as observational until a randomized test says otherwise.

An average by calendar is a photograph taken with a shaky hand. Before you plan around a gentle slope, check that it is not a cliff seen through the blur.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk