HRT and Alzheimer's risk: what the 2026 study found, and what it did not

What the headline will say, and what the study tested

A result like this travels as "HRT protects against Alzheimer's." The paper does not say that, and the gap between the two is where a reader gets misled. Here is each claim you are likely to encounter set against what the cohort actually measured.

What did the 2026 Neurology study actually measure?

Two independent data sets, with different outcomes in each. Neuropathology came from the National Alzheimer's Coordinating Center, comparing 258 women who reported estrogen-only HRT use (mean age at death 81.9) against 2,701 who reported no use (mean age at death 82.2). Biomarker and clinical outcomes came from the Alzheimer's Disease Neuroimaging Initiative, comparing 110 users (mean age 76.5) against 1,948 non-users (mean age 73.2).

The primary outcome was the odds of increased Alzheimer pathology on autopsy. The reported figures:

• Increased Alzheimer pathology at autopsy: OR 0.65 (95% CI 0.48-0.88, p = 0.005)
• Amyloid load measured in plasma: β = 0.44 (95% CI 0.16-0.73, p = 0.0025)
• Amyloid load measured in cerebrospinal fluid: β = 0.07 (95% CI 0.002-0.13, p = 0.030)
• Clinical dementia diagnosis: OR 0.61(95% CI 0.55-0.67, p < 0.0001)
• Symptoms of memory or functional decline: OR 0.67(95% CI 0.61-0.74, p < 0.0001)

Source: Bruno JJ, Shaw JS, Hosseini SMH, for the Alzheimer's Disease Neuroimaging Initiative, Association Between Menopausal Hormone Therapy and Alzheimer Disease Neuropathology, Neurology 2026;107(5):e218413, doi 10.1212/WNL.0000000000218413, epub August 12, 2026.

Why an autopsy cohort cannot answer the causal question

Participants reported their own HRT use, and the researchers looked at what happened afterwards. Nobody was assigned to take estrogen or not. Women who took HRT in that era differed from women who did not in ways that also track with dementia risk, including access to care, education, and general health at the point of prescription. That is the standard limitation of observational work, and it is why the authors write that their results do not address causality and cite limited generalizability from the retrospective design.

The accompanying editorial in the same issue frames the open questions as "When, Which, and for Whom?" (Salemme S, Neurology 2026;107(5):e218503, doi 10.1212/WNL.0000000000218503). Timing, formulation, and patient selection are unresolved, which is a fair reading of where the evidence sits.

What did the UK Biobank cohort find?

Fourteen days after the autopsy paper, a much larger study landed in Alzheimer's & Dementia. Researchers at the University of East Anglia and the University of Exeter followed 183,450 postmenopausal women in UK Biobank for a mean of 13.3 years, covering 2.43 million person-years, and identified 3,948 dementia cases. The exposure was HRT use for at least one year.

The single most useful thing in the paper is that it reports three endpoints separately, and they do not agree with each other. Press coverage tends to quote one and name another.

By menopause type, the split is sharper still. For women who had a surgical menopause, meaning a hysterectomy or removal of both ovaries, HRT use was associated with a hazard ratio of 0.74 (95% CI 0.65-0.84) for dementia of any cause and 0.68 (95% CI 0.56-0.82) for Alzheimer's. For women who reached menopause naturally, the all-cause figure was 0.94 (95% CI 0.87-1.01), which did not reach significance. The interaction test for menopause type came out at p = 0.001, so this is a stratification the data supports rather than a slice that happened to look interesting.

The authors apply their own brake to another of the subgroups. The APOE4 finding, a hazard ratio of 0.87 (95% CI 0.80-0.95) in carriers for dementia of any cause, comes with their note that no significant overall interaction between APOE4 and HRT was observed and that it "should be interpreted with a degree of caution". The section below sets all four subgroups against the test each one did or did not get.

One detail worth knowing, because the press releases smooth it over by saying "compared with those who had never used it". The comparison group is not purely never-users. Per the paper, the reference is "those who never took HRT or took HRT but started and stopped at the same age", which folded in 9,065 women who used HRT for under a year. The authors ran the analysis again with those women removed and reported that the results did not change significantly.

Source: Squires S, Saleh RN, Pilling LC, Atkins JL, Ranson JM, Tai XY, Vauzour D, Llewellyn DJ, Minihane AM, Hormone replacement therapy and dementia risk among postmenopausal women: Identifying responsive subgroups in the UK Biobank, Alzheimer's & Dementia 2026;22(8):e71679, doi 10.1002/alz.71679, published August 26, 2026.

Which subgroups responded, and which the authors tested

The paper is titled "Identifying responsive subgroups in the UK Biobank", so the subgroups are the argument rather than a footnote to it. There are four of them, and they do not carry equal weight. Two arrive with a significant interaction test. One arrives with a warning from the authors. The fourth, which is the one that produced every headline about a critical window, arrives with no test reported at all.

Read from sections 3.2 and 3.5 of the full text at PMC. Every figure here is for dementia of any cause, so that the four rows can be compared with each other.

Were the subgroups prespecified? The paper never uses that word, and it never uses "post hoc" either, so the useful thing is to report what it does say. Its Methods section names APOE4 status, surgical menopause and length of natural estrogen exposure as "possible modifying effects" the team set out to investigate. The age-at-initiation analysis is declared one paragraph later and framed differently, as work done "to explore the window of opportunity hypothesis", which is a named idea that predates this data by more than a decade. All four are declared ahead of the results, and identifying responsive subgroups is the stated purpose in the title. What the statistical analysis section does not describe anywhere is an adjustment for multiple comparisons.

One phrase in the paper is worth reading closely, because it is the version that travels. The abstract says initiation between ages 46 and 56 "conferred the greatest benefit". Conferred is a causal verb. The limitations section of the same paper says "results are associations and not causal". Both sentences belong to the authors. The second one governs, and it is the one that drops out of the coverage.

Does the age HRT was started change the association?

This is the part the order of publication makes easy to miss, and it is the part with the most practical content. The UK Biobank authors broke their cohort down by the age at which women first took HRT, and the association moves a long way depending on where you land.

Hazard ratios for dementia of any cause, with 95% confidence intervals and the p values the paper prints, read from section 3.5. The surgical under-46 band is written "p < 0.04" in the paper and is reproduced here as the authors wrote it. Cells marked "not reported in the text" are rendered in the paper's Figure 4 but are not given as numbers in the prose, so we do not print numbers for them. The paper also states that starting earlier than 47 or later than 56 was associated with a non-significant increase in dementia risk, in a sentence that sits between its natural-menopause and surgical-menopause results without naming which group it applies to. We leave that attribution open rather than pick one.

The authors read this as support for what they call the window of opportunity hypothesis, the older idea that hormone therapy does something different to a brain near the menopausal transition than to a brain fifteen years past it. Press coverage generally calls the same idea the critical window.

How the window was drawn matters for how much weight it holds. The authors did not simply cut the cohort into four bands. Per their Methods, they swept a threshold across every starting age from 40 to 59, asking at each one what the hazard looked like for women who began HRT before that age and then for women who began after it. That sweep is the continuous curve in the paper's Figure 4. The four-band table above is a coarser second rendering of the same underlying data.

Which is where the missing test matters. For age at first use the paper reports no interaction term and no test for trend across the bands. It reports both for menopause type, at p = 0.001, and for length of estrogen exposure, at p = 0.018, so the gap stands out against the paper's own practice rather than against an outside standard. The p values in the table are each band measured against women who never used HRT. A p value there tells you that band differed from never-users. It says nothing about whether the bands differ from each other, and the bands differing from each other is the entire content of the claim that timing matters.

Two more things the main text withholds. The timing analysis was repeated for the Alzheimer and non-Alzheimer endpoints separately and those results were placed in the supplement, so the paper's prose carries no hazard ratio for Alzheimer's by starting age. And the reverse-causation check the authors ran, removing women diagnosed with dementia within 2, 4, 6 and 8 years of baseline, is reported against the overall HRT association; the main text does not report it repeated for the age bands.

It is a hypothesis, and it should be held as one. A cohort study can show that women who started HRT at 53 had different outcomes from women who started at 62, and it cannot show why. The women who start later are a different group of women, arriving with different reasons for the prescription. The authors say this in their own discussion: at older ages they "cannot preclude the possibility that reverse causation or confounding by indication affected the associations". They close by calling for confirmation in a randomised trial.

Where the two 2026 studies agree, and where they part company

They point the same direction. Neither of them can prove the direction is real, and they are weak and strong in opposite places, which is the useful thing about having both.

The real disagreement in this field is not between these two papers. It is between every observational study of HRT and the one randomised trial that used dementia as an outcome. As the UK Biobank authors put it in their own discussion, WHIMS "observed an increased incidence of dementia over 5 years following estrogen plus progestogen HRT use in participants > 65 years". That trial found harm. These cohorts find association with benefit. The window of opportunity hypothesis is the proposed reconciliation, since WHIMS enrolled women well past menopause, and it is proposed rather than demonstrated.

The UK Biobank paper carries limitations worth stating plainly, because they are the authors' own. Women who take HRT differ systematically from women who do not, in socioeconomic status, education, ethnicity and lifestyle, and while those were adjusted for, the authors write that they cannot rule out residual confounding explaining part of the 10 percent overall reduction. UK Biobank participants are predominantly of European descent and skew more socioeconomically advantaged than the general population. Dementia diagnoses came mainly from hospital inpatient records, which miss milder cases. And the paper reports that when stratifying into subgroups, statistical power to draw strong conclusions is limited.

How this sits against the FDA label

In February 2026 the FDA approved labeling changes removing the boxed warning from six menopausal hormone therapy products, deleting the cardiovascular-disease, breast-cancer, and probable-dementia statements. (FDA) Two separate things follow from that, and they are easy to run together.

Removing the probable-dementia statement corrected an overstatement of harm drawn largely from a 2002 trial of much older women. It did not add a benefit claim. The labeling still states these products are not approved to prevent dementia, so a clinician has no label basis for prescribing HRT as a cognitive prophylactic. The full account of what came off the label and what remains is in our black box warning guide.

So should you start HRT for brain protection?

Not on this evidence, and two observational studies pointing the same way do not add up to one trial. The established reasons to take HRT are menopausal symptoms and bone protection, and the timing rule that governs the risk-benefit balance is about starting within roughly ten years of menopause. What the UK Biobank cohort adds is that the same rough window shows up again when the outcome is dementia rather than heart or bone, which is suggestive and is not a licence. If you are weighing HRT for symptoms and these findings are one input among several, that is a reasonable way to hold them. If they are the reason, neither study carries that weight.

What to do with it in a consultation: ask which formulation is being proposed, since the autopsy cohort tested estrogen-only therapy and the UK Biobank cohort recorded no formulation at all, and ask how your own timing and history change the balance. If you had a surgical menopause, that is worth raising specifically, because it is the subgroup where the UK Biobank cohort found its largest associations, on both of its endpoints. The Neurology cohort reports no menopause-type subgroup. Our 2026 safety guide covers the risk picture in full, and the perimenopause hub covers starting decisions for the younger cohort.

Common questions

Sources

• Bruno JJ, Shaw JS, Hosseini SMH, for the Alzheimer's Disease Neuroimaging Initiative. Association Between Menopausal Hormone Therapy and Alzheimer Disease Neuropathology. Neurology 2026;107(5):e218413. doi 10.1212/WNL.0000000000218413. PubMed 42585606
• Salemme S. Menopausal Hormone Therapy and Alzheimer Disease: When, Which, and for Whom? Neurology 2026;107(5):e218503. doi 10.1212/WNL.0000000000218503. PubMed 42585609
• Squires S, Saleh RN, Pilling LC, Atkins JL, Ranson JM, Tai XY, Vauzour D, Llewellyn DJ, Minihane AM. Hormone replacement therapy and dementia risk among postmenopausal women: Identifying responsive subgroups in the UK Biobank. Alzheimer's & Dementia 2026;22(8):e71679. doi 10.1002/alz.71679. Full text at PMC · PubMed 42644405
• University of Exeter. HRT linked to lower dementia risk in large UK study. news.exeter.ac.uk
• FDA. FDA Approves Labeling Changes for Menopausal Hormone Therapy Products. fda.gov