How Countries Count Dementia, and Why Their Figures Don't Line Up

There is no single way the world counts dementia. Countries use GP registers, hospital records, insurance claims, household surveys and research cohorts, and each one sees a different slice of the people living with the condition. A diagnosis rate in one country and a prevalence estimate in another are often not measuring the same thing, and should not be set side by side as if they were.

Behind every national dementia figure are people: someone whose diagnosis was written into a GP record, someone assessed by a researcher at home, someone who has never seen a doctor about their memory and so appears in no register at all. How a country chooses to count decides which of them are visible. What is counted is what gets planned for, funded and improved. This article sets out the main methods, what each finds and misses, and why their figures rarely line up across borders.

Change the definition of dementia and the count can change tenfold

Before any system can count dementia, someone must decide what counts as dementia, and that is not settled. In a 1997 study, researchers applied six widely used sets of diagnostic criteria to the same 1,879 people aged 65 and over in the Canadian Study of Health and Aging. The share classed as having dementia ran from 3.1% under the World Health Organization’s ICD-10 criteria to 29.1% under the American Psychiatric Association’s older DSM-III criteria. Only 20 people met all six definitions. The authors concluded that commonly used criteria “can differ by a factor of 10” in the number of people classed as having dementia, and warned that this matters for older people’s rights “to drive, make a will, and handle financial affairs”^[1]. The criteria have since been revised, but the lesson stands: a prevalence figure is only as comparable as the definition behind it.

Definitions also travel badly across cultures and levels of schooling. The 10/66 Dementia Research Group surveyed all residents aged 65 and over (14,960 people) in 11 sites across seven low- and middle-income countries: China, India, Cuba, the Dominican Republic, Venezuela, Mexico and Peru. Using a computerised version of the DSM-IV criterion, dementia prevalence ranged from 0.3% in rural India to 6.3% in Cuba. Using the 10/66 algorithm, built to be fair to people with little formal education, it was higher and more consistent, from 5.6% in rural China to 11.7% in the Dominican Republic^[4]. The 847 people classed as having dementia by 10/66 but not by DSM-IV had high levels of disability, which supports the 10/66 diagnosis. The researchers concluded that DSM-IV “might underestimate dementia prevalence, especially in regions with low awareness”^[4]. Some of what once looked like lower dementia rates in poorer countries may have been partly a product of the yardstick.

Five ways of counting, and what each one misses

Once a definition is chosen, countries gather figures in broadly five ways.

The first three count diagnoses: they find only people a health service has already recognised. Only surveys look for dementia itself; models fill the gaps with arithmetic.

Many countries have little of any of this. WHO’s Global Dementia Observatory collates data from Member States on 35 key dementia indicators; 62 countries contributed in its first round of data collection, 2018 to 2020^[5]. Among the countries taking part, only one in three routinely monitor people with dementia, and fewer than half monitor services for them. WHO’s global dementia action plan set a target of 50% of countries routinely collecting a core set of dementia indicators every two years by 2025^[3]. How far countries remain from that target is covered in our article on dementia diagnosis rates.

Registers count diagnoses, not dementia, and each register counts differently

England offers a worked case of a primary care register, because its data are unusually complete. Recording is mandatory for GP practices, and in July 2025 the national dataset covered 98.67% of practices (6,138 of 6,221)^[6]. On 30 June 2026, 516,230 people had a recorded dementia diagnosis. NHS England’s headline diagnosis rate, 66.2% at that date, is not a count of anything directly: it is the number of people aged 65 and over with a recorded diagnosis, divided by the number “estimated to have dementia”^[7]. That denominator is modelled from the Cognitive Function and Ageing Study II (CFAS II), which, a 2026 review notes, was “conducted over 15 years ago” and is not adjusted for ethnicity, deprivation or rurality. As a result, regional differences in the rate “may reflect differences in population structure” rather than real differences in how well areas diagnose^[6]. None of the other registries the review compared (in Sweden, Norway and Australia) routinely publishes a diagnosis rate of this kind^[6]. Even within England, the current series cannot be compared with data before October 2022, because of retrospective coding, changes in patient registration and differences in coverage and definitions^[7].

The Nordic and Australian registries work differently. Sweden’s SveDem, set up in 2007, gathers diagnoses from 57 specialist units and 926 primary care units, and patients opt in. Norway’s NorCog draws on geriatric, old-age psychiatry and memory clinics and captured about 70% of dementia diagnoses in 2024, compared with the national patient registry. Australia’s ADNeT registry covers 82 memory clinic sites^[6]. These are quality registries, designed to check and improve care for the people in them, not to count everyone with dementia.

Registers can also be accurate about the syndrome and weak on the detail. When researchers checked 197 records from Denmark’s nationwide hospital registers, a registered dementia diagnosis was correct in 169 (85.8%). But agreement on the type of dementia was low (kappa 0.36), and Alzheimer’s disease was under-registered^[8]. The share of diagnoses with no type recorded at all also varies, as the chart shows.

Claims and administrative data find diagnosed dementia well, and early dementia poorly

Where there is no dementia register, many countries count people through the records health systems already keep for paying bills. Canada’s national surveillance system is a clear example. It counts a person aged 65 or over as having diagnosed dementia if they have at least one hospital record or at least three physician claims within two years, at least 30 days apart, carrying a dementia code, or at least one prescription for an anti-dementia drug. On that definition, more than 402,000 older people were living with diagnosed dementia in Canada in 2013–14, excluding Saskatchewan, whose data were not included^[9]. The agency notes that such figures can shift with “changes in data collection methods, coding and classification systems”, and that “dementia estimates vary among population-based studies”^[9].

How good are these records? In the US, researchers compared Medicare claims with in-person clinical assessments of 758 people in the Aging, Demographics and Memory Study. Claims picked up 85% of people with dementia (sensitivity) and correctly left out 89% of those without it (specificity); for Alzheimer’s disease, sensitivity fell to 64%. Because of the false positives, claims “overcount the true prevalence of dementia” in that sample, while still missing some people^[10]. An earlier US study of 417 people with clinically diagnosed Alzheimer’s disease found that five years of claims identified 87% as having dementia, hospital files alone identified only 29% as having Alzheimer’s disease, and people with mild dementia were less likely to be identified at all^[11].

Across countries, the picture is wide. A 2018 systematic review of 27 studies comparing coded routine health data with expert diagnosis found that, for all-cause dementia, the share of coded cases that were genuine (positive predictive value) ranged from 33% to 100%, though 16 of 27 studies were above 75%. Sensitivity ranged from 21% to 86%. Positive predictive values were generally lower for vascular dementia (19% to 91%) than for Alzheimer’s disease (57% to 100%)^[2]. In short, routine records are often right about the people they flag, but how many people they miss depends heavily on the setting, and the authors advise each study to run its own setting-specific validation^[2].

Surveys and cohorts look for dementia directly, but each uses its own yardstick

The only way to count people whose dementia has never been diagnosed is to go out and assess a representative sample. In the US, the Health and Retirement Study’s Harmonized Cognitive Assessment Protocol gave 3,496 people aged 65 and over a full battery of cognitive tests and an informant interview in 2016. Using an algorithm based on standard diagnostic criteria, it classed 10% as having dementia and 22% as having mild cognitive impairment, weighted to the national population^[12]. It replaced a diagnostic substudy that “was fielded more than 20 years ago”^[12].

Because in-depth assessment is costly, many studies instead classify dementia with an algorithm from the shorter tests in a large survey. These algorithms disagree. Five used with the US Health and Retirement Study were compared against clinical assessment: in the data they were built on, sensitivity ranged from 53% to 90%; in separate validation data, it fell to between 18% and 62%. Accuracy was higher among younger, highly educated and non-Hispanic white participants, and the authors warn that naive use of these algorithms in research on disparities may bias results^[13]. An algorithm that works less well in some communities can distort comparisons between them.

India shows how much a well-designed survey can add. Earlier estimates came from “samples from selected communities”. The Longitudinal Aging Study in India obtained a clinical consensus rating for 2,528 people, then used a statistical model to estimate dementia status for 28,949 respondents aged 60 and over, surveyed 2018–2020. It put prevalence at 7.4%, about 8.8 million people^[14].

Global totals rest on a further layer of modelling. The Global Burden of Disease study estimated 57.4 million people with dementia worldwide in 2019 (uncertainty interval 50.4 to 65.1 million), and forecast 152.8 million in 2050, using regression models that include education and three risk factors^[15]. How these projections are made is covered in our article on dementia prevalence and projections.

What the evidence asks of us

Better counting is not an end in itself. People living with dementia who are missing from the data are missing from the plans. The evidence points to five things health systems, statistics agencies, funders and researchers can do.

• Meet the global monitoring target. WHO’s target was for half of countries to collect a core set of dementia indicators every two years by 2025; among Global Dementia Observatory countries, only one in three routinely monitor people with dementia^[3]. Governments that do not yet report should start, and those that do should report regularly.
• Pair every register with a population survey. Registers and claims count only people already diagnosed, and in US claims, people with mild dementia were less likely to be identified^[11]. Assessment studies such as the US Health and Retirement Study’s cognitive protocol and the Longitudinal Aging Study in India show that representative national measurement is possible^[12][14], and they should be repeated so that estimates do not drift decades out of date.
• Keep denominators current and fair. Where a diagnosis rate depends on a modelled estimate, the model should be updated and should reflect differences in population, such as ethnicity, deprivation and rurality, which England’s current estimate does not^[6].
• Check that records are right. Accuracy of routine data varies widely between settings, so studies and agencies relying on them should validate them locally^[2], and record the type of dementia, not just the syndrome^[6][8].
• Use yardsticks that are fair to everyone. Criteria and algorithms that work less well for people with little schooling, or for some ethnic groups, can misclassify them^[4][13]. Methods validated across cultures and education levels, as the 10/66 algorithm was designed to be^[4], should be the norm.

Anyone comparing countries should say what each figure measures. A recorded diagnosis, a claims count and a survey estimate are different things; treating them as one hides the people the figures are meant to serve.

Language: this piece uses person-first language ("people living with dementia", "people with dementia"), as most dementia organisations and people with lived experience prefer, and uses clinical terms such as "Alzheimer's disease" or "mild cognitive impairment" only where a study, dataset or diagnosis refers to that specific condition.

Why we're publishing this

Dementia charities, carers’ organisations and support groups are often asked to make their case with numbers, and are just as often told that the numbers don’t agree. We compiled this evidence so that they can explain why: different countries are counting different things. The case for better data belongs to people living with dementia and the people who support them, because those left out of the figures are the ones most likely to be left out of services. This article is part of our Dementia & Health research library; for how countries compare on what the figures do show, see our international comparison.

How to read this data

Most figures here describe how well a way of counting works, not how many people have dementia. Sensitivity is the share of people with dementia that a method finds; specificity is the share of people without dementia it correctly leaves out; positive predictive value is the share of people it flags who really have dementia. The validation studies are from high-income countries and some are old (the criteria study is from 1997, the Danish and Medicare validations from 2002–2009); coding and criteria have changed since. England’s figures are recorded diagnoses set against a modelled estimate from CFAS II; the Canadian figure is administrative, from 2013–14. The US and Indian prevalence figures come from assessment of representative samples using algorithms, and the global total is a model. The registry figures come from each registry’s annual report as compiled in a 2026 review, with different coverage and denominators. None should be compared directly across countries.

Use this data

Free to cite with attribution to Understanding Aging. Charities, carers' groups, campaigners and journalists are welcome to reuse these figures and charts. For the underlying figures as a spreadsheet, get in touch.

Important: This article summarises published research and statistics for general information. It is not medical, care or financial advice and cannot be used to diagnose, treat or predict anything about an individual. If you have concerns about your own or someone else's memory or health, please speak to a GP or another qualified health professional.

Sources

1. Erkinjuntti T, Ostbye T, Steenhuis R, Hachinski V. The effect of different diagnostic criteria on the prevalence of dementia. New England Journal of Medicine, 1997. europepmc.org/article/MED/9385127
2. Wilkinson T, Ly A, Schnier C, et al. Identifying dementia cases with routinely collected health data: a systematic review. Alzheimer’s & Dementia, 2018. europepmc.org/article/MED/29621480
3. World Health Organization. Global Dementia Observatory: information systems. WHO, accessed 2026. who.int/data/gho/data/themes/global-dementia-observatory-gdo/information-systems
4. Llibre Rodriguez JJ, Ferri CP, Acosta D, et al. Prevalence of dementia in Latin America, India, and China: a population-based cross-sectional survey. The Lancet, 2008. europepmc.org/article/MED/18657855
5. World Health Organization. Global Dementia Observatory (GDO). WHO, accessed 2026. who.int/data/gho/data/themes/global-dementia-observatory-gdo
6. Hazan J, Howard R, Isaacs JD. England’s dementia diagnosis landscape: strengths, limitations, and future directions. International Journal of Geriatric Psychiatry, 2026. europepmc.org/article/PMC/PMC13327468
7. NHS England. Primary Care Dementia Data, June 2026. NHS England Digital, 2026. digital.nhs.uk/data-and-information/publications/statistical/primary-care-dementia-data/june-2026
8. Phung TK, Andersen BB, Høgh P, et al. Validity of dementia diagnoses in the Danish hospital registers. Dementia and Geriatric Cognitive Disorders, 2007. europepmc.org/article/MED/17690555
9. Public Health Agency of Canada. Dementia in Canada, including Alzheimer’s disease: highlights from the Canadian Chronic Disease Surveillance System. Government of Canada, 2017. canada.ca/en/public-health/services/publications/diseases-conditions/dementia-highlights-canadian-chronic-disease-surveillance.html
10. Taylor DH, Østbye T, Langa KM, Weir D, Plassman BL. The accuracy of Medicare claims as an epidemiological tool: the case of dementia revisited. Journal of Alzheimer’s Disease, 2009. europepmc.org/article/MED/19542620
11. Taylor DH, Fillenbaum GG, Ezell ME. The accuracy of Medicare claims data in identifying Alzheimer’s disease. Journal of Clinical Epidemiology, 2002. europepmc.org/article/MED/12393082
12. Manly JJ, Jones RN, Langa KM, et al. Estimating the prevalence of dementia and mild cognitive impairment in the US: the 2016 Health and Retirement Study Harmonized Cognitive Assessment Protocol Project. JAMA Neurology, 2022. europepmc.org/article/MED/36279130
13. Gianattasio KZ, Wu Q, Glymour MM, Power MC. Comparison of methods for algorithmic classification of dementia status in the Health and Retirement Study. Epidemiology, 2019. europepmc.org/article/MED/30461528
14. Lee J, Meijer E, Langa KM, et al. Prevalence of dementia in India: national and state estimates from a nationwide study. Alzheimer’s & Dementia, 2023. europepmc.org/article/MED/36637034
15. GBD 2019 Dementia Forecasting Collaborators. Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study 2019. The Lancet Public Health, 2022. europepmc.org/article/MED/34998485