Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study 2019

1A growing body of evidence from North America and Europe suggests a decreasing trend in dementia incidence, potentially due to increases in educational attainment and improvements in the management of cardiovascular disease and its risk factors.,, It remains unclear whether these trends will continue into the future or whether they extend to other geographical areas, with previous work in Japan and China suggesting increases in age-specific prevalence, which is a function of disease incidence and duration., However, all estimates agree that the absolute number of people affected by dementia will show large increases over time.,,, Globally, the number of people affected by dementia was estimated to have increased by 117% (95% uncertainty interval [UI] 114–121) between 1990 and 2016, largely due to population ageing.

2Demographic analyses suggest that these patterns are driven by decreases in fertility coupled with increases in life expectancy, which together lead to large changes in the age structure of the population (larger numbers of people at the oldest ages than was the case historically)., These changes, along with largely stable age-specific prevalence estimates and population growth, lead to large increases in the number of people affected by dementia. Given that these demographic trends are expected to continue into the future, the number of people with dementia will continue to rise.

3Evidence before this study Global efforts to summarise the prevalence and burden of dementia, including the World Dementia Report and the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD), can serve to guide resource allocation and health policy decision making. Additionally, projections of future dementia burden are essential to guide health system planning and inform research funding decisions. We searched PubMed on Oct 23, 2020, using the search terms [“dementia” OR “alzheim*”] AND [“forecast” OR “project*”] AND “prevalence.” Of 635 articles published in English, three reported on global estimates of projected prevalence for dementia or Alzheimer's disease and 30 reported on country-specific estimates. The three studies that calculated global estimates either assumed a constant prevalence and applied population estimates to calculate the effect of changing demographics, or used macrosimulation techniques to estimate prevalence on the basis of transition probabilities for incidence, progression of disease, and mortality. However, none of the available global studies included comprehensive country-level estimates, nor did they consider the potential effects of trends in exposure to known dementia risk factors. Added value of this study This study leveraged country-specific estimates of dementia prevalence from the GBD study to project dementia prevalence globally, by world region, and at the country level. Furthermore, we incorporated information on projected trends in exposure to known dementia risk factors to understand how trends in risk factors might affect the projected number of individuals with dementia, and did decomposition analysis to understand the drivers of forecasted changes. Our estimates of age-standardised prevalence remained stable between 2019 and 2050 (percentage change of 0·1% [95% uncertainty interval −7·5 to 10·8]), whereas the number of individuals estimated to have dementia increased considerably, from 57·4 (50·4–65·1) million cases in 2019 to 152·8 (130·8–175·9) million cases in 2050. Although we projected increases in the estimated number of individuals with dementia in every country, there was considerable variation, with the highest projected increases in regions of north Africa and the Middle East, and eastern sub-Saharan Africa. Implications of all the available evidence Due to increases in population growth and population ageing, huge increases in the number of individuals affected by dementia can be expected in 2050. Given that there are currently no available disease-modifying therapies, appropriate emphasis should be placed on efforts to address known modifiable risk factors. Multimodal interventions have shown some success in delaying the rate of cognitive decline and present a promising approach to risk reduction and dementia prevention. Simultaneously, it will be necessary to plan for the expected increases in the utilisation of health and social care services and to expand resources to support caregivers of individuals with dementia. Finally, continued resources should be directed towards better understanding and characterising disease mechanisms, with the goal of developing effective therapeutic agents.

4Evidence before this study

5Global efforts to summarise the prevalence and burden of dementia, including the World Dementia Report and the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD), can serve to guide resource allocation and health policy decision making. Additionally, projections of future dementia burden are essential to guide health system planning and inform research funding decisions. We searched PubMed on Oct 23, 2020, using the search terms [“dementia” OR “alzheim*”] AND [“forecast” OR “project*”] AND “prevalence.” Of 635 articles published in English, three reported on global estimates of projected prevalence for dementia or Alzheimer's disease and 30 reported on country-specific estimates. The three studies that calculated global estimates either assumed a constant prevalence and applied population estimates to calculate the effect of changing demographics, or used macrosimulation techniques to estimate prevalence on the basis of transition probabilities for incidence, progression of disease, and mortality. However, none of the available global studies included comprehensive country-level estimates, nor did they consider the potential effects of trends in exposure to known dementia risk factors.

6Added value of this study

7This study leveraged country-specific estimates of dementia prevalence from the GBD study to project dementia prevalence globally, by world region, and at the country level. Furthermore, we incorporated information on projected trends in exposure to known dementia risk factors to understand how trends in risk factors might affect the projected number of individuals with dementia, and did decomposition analysis to understand the drivers of forecasted changes. Our estimates of age-standardised prevalence remained stable between 2019 and 2050 (percentage change of 0·1% [95% uncertainty interval −7·5 to 10·8]), whereas the number of individuals estimated to have dementia increased considerably, from 57·4 (50·4–65·1) million cases in 2019 to 152·8 (130·8–175·9) million cases in 2050. Although we projected increases in the estimated number of individuals with dementia in every country, there was considerable variation, with the highest projected increases in regions of north Africa and the Middle East, and eastern sub-Saharan Africa.

8Implications of all the available evidence

9Due to increases in population growth and population ageing, huge increases in the number of individuals affected by dementia can be expected in 2050. Given that there are currently no available disease-modifying therapies, appropriate emphasis should be placed on efforts to address known modifiable risk factors. Multimodal interventions have shown some success in delaying the rate of cognitive decline and present a promising approach to risk reduction and dementia prevention. Simultaneously, it will be necessary to plan for the expected increases in the utilisation of health and social care services and to expand resources to support caregivers of individuals with dementia. Finally, continued resources should be directed towards better understanding and characterising disease mechanisms, with the goal of developing effective therapeutic agents.

10Additionally, evidence has emerged supporting the importance of potentially modifiable risk factors for dementia. The 2020 update to the Lancet Commission on dementia prevention, intervention, and care highlighted the evidence for 12 modifiable risk factors for dementia: low education, hypertension, hearing impairment, smoking, midlife obesity, depression, physical inactivity, diabetes, social isolation, excessive alcohol consumption, head injury, and air pollution. Addressing these factors through public health interventions is a pathway towards reducing disease prevalence, and future changes in modifiable risk factors might influence the trajectory of trends in age-specific prevalence. Studies have hypothesised that education and the prevalence and treatment of cardiovascular disease might be key in explaining previous and future trends in dementia prevalence., Therefore, incorporating the potential effect of changes in risk factors should be a key component of any forecast of dementia prevalence.

11The increases over time in the number of people affected by dementia accentuate the importance of not only quantifying the current burden of dementia, but also generating forecasts of the prevalence of dementia to allow for informed policy decisions, health system planning, and allocation of resources. Previous efforts to forecast the prevalence of dementia have largely either been country-specific or have predicted future global cases by simply applying global prevalence estimates to population projections.,,, Other work has applied macrosimulation techniques, using estimates of incidence, progression, and mortality, although the data requirements to fit these models are large while the evidence base for such data is sparse.

12This Article aims firstly to improve and expand on previous forecasting estimates by incorporating the estimated effect of future changes in selected modifiable risk factors and projecting dementia prevalence until 2050 globally, as well as by world region and country, and secondly to decompose the drivers of changes in dementia prevalence. This Article was produced as part of the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) Collaborator Network and in accordance with the GBD Protocol.

13The category of dementia, as described in this Article, is referred to under the name Alzheimer's disease and other dementias in GBD 2019. General GBD methods can be found in the GBD 2019 summary papers and in the appendix (pp 10–11).,, The overall GBD protocol and data visualisation tools can be found online. This study complies with the Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER), and a completed GATHER checklist is available in the appendix (pp 3–4). Flow charts summarising the methodology for the estimation of dementia prevalence from 1990 to 2019 and the projection of dementia prevalence to 2050 are available in the appendix (pp 5–6).

14The reference case definitions for the estimation of dementia prevalence in the GBD studies were from the Diagnostic and Statistical Manual of Mental Disorders (DSM)-III, DSM-III-R, DSM-IV, or DSM-5, or the International Classification of Diseases (ICD)-8, ICD-9, or ICD-10, assigned by study clinicians in representative surveys., Additional data collected using different definitions or ascertainment methods were included and were adjusted to account for bias due to differences in study characteristics. In GBD 2019, the prevalence of dementia was adjusted downwards to account for dementia that is caused by other clinical diseases (eg, clinical stroke, Parkinson's disease, Down syndrome, and traumatic brain injury). This correction ensures that the GBD cause list is mutually exclusive. However, dementia as described in this Article does not include this correction, because the future total burden of all dementias is of policy relevance and has been the focus of previous forecasting efforts. Therefore, estimates presented in this Article will not align with those presented in the online tools. The data sources used in this study were identified through systematic review (appendix pp 12–13) and can be downloaded from the Global Health Data Exchange.

15We used a Bayesian meta-regression model, MR-BRT (Meta-Regression–Bayesian Regularised Trimmed), to estimate the sex ratio from sex-specific data and applied this ratio to split data only reported for both sexes combined (34 studies). Additionally, we age-split data with an age range greater than 20 years using the global age pattern, estimated from all data available in narrow age bins (61 studies). To adjust for bias by study definition or characteristics, we used MR-BRT to estimate a network meta-regression model of the bias between the reference case definition and alternative criteria. This modelling framework allows for the inclusion of evidence on the association between different alternative methods rather than limiting the analysis to data on direct comparisons between the reference criteria and each alternative criterion. We modelled the bias associated with studies that used clinical records for diagnosis, studies that used algorithms to classify dementia cases, studies that applied the National Institute on Aging–Alzheimer's Association diagnostic criteria, studies that used general practice records (eg, electronic medical records for diagnosis), and studies that used the 10/66 diagnostic algorithm., We also adjusted for bias in US medical claims data but, because there was evidence of differential bias by age, we used a separate model with a spline on age to appropriately make this adjustment (appendix p 14).

16We used Disease Modelling Meta-Regression (DisMod-MR) 2.1, a Bayesian compartmental model widely used in GBD non-fatal modelling, to generate initial estimates of dementia prevalence. This model ensures consistency between the different parameters (ie, prevalence, incidence, remission, mortality) by enforcing associations in a set of differential equations. We assumed zero remission and no incidence before 40 years of age given that dementia is a progressive, chronic disease and is extremely rare before this age., The Bayesian modelling framework allows for the estimation of prevalence in locations with little or no data by leveraging information from covariates as well as priors based on the model fit from the previous higher level of the location hierarchy (ie, the prior for a specific country would be based on the model fit for the region). In locations with high-quality data for both prevalence and incidence, we found that the data on incidence combined with mortality estimates implied a higher prevalence than that observed in the data. Given that dementia incidence is probably measured with less precision because of the gradual onset of disease, we chose to retain prevalence data in our models and exclude data on incidence to improve the estimation of prevalence in this model. Although the measurement of prevalent dementia is not without error, because of the higher average disease severity among prevalent cases than among incident cases, measurement error is probably lower among prevalent cases. In this model, we used covariates of smoking prevalence and average years of education in the population aged 15 years and older to help to guide the model estimation in locations with poor-quality data or no data. We set priors on the value of the education parameter on the basis of the results of a literature meta-analysis on the relative risk of dementia for each additional year of education (appendix p 15). This prevalence model used literature data on mortality to guide initial estimates of dementia mortality, and estimates from this model served as inputs to the mortality modelling process (appendix pp 16–19).

17After estimating mortality due to dementia, we ran a second DisMod-MR 2.1 model to calculate final prevalence estimates. This model included the same settings as the first model but excluded literature data on mortality and included our GBD dementia mortality estimates.

18We initially included the three risk factors (high body-mass index [BMI], high fasting plasma glucose, and smoking) for which sufficient evidence exists to merit inclusion as risk factors for dementia in GBD 2019.,,, However, various other risk factors have also been proposed for dementia and we tested the association between dementia prevalence and risk factor prevalence for other risks that were included in the 2020 Lancet Commission report and are quantified within the GBD framework (low physical activity, high systolic blood pressure, low education, alcohol use, and air pollution). To quantify risk factors other than education, we used the summary exposure value (SEV), which is a risk-weighted prevalence of a given risk factor exposure, first developed for GBD 2015. The SEVs range from 0 to 1, with a value of 0 indicating that there is no risk (or no protection for protective risk factors) in a population and a value of 1 indicating that the entire population is at maximum risk.

19To forecast SEVs from 2019 to 2050, we calculated the annualised rate of change in logit SEV for each location, age, sex, and past year. We then estimated future annualised rates of change by calculating a weighted mean of prior annualised rates over the time series. A recency weighting parameter altered how much weight is given to more recent trends versus older trends. This parameter was selected using out-of-sample predictive validity holding out the last 10 years of data available. Forecasts for this period were then used to calculate the root-mean square error, which informed the selection of weights. To reduce the effect of extreme trends in long-range forecasts, we implemented caps at the 5th and 95th percentiles of logit SEVs.

20We forecasted years of education (up to a maximum of 18 years of education) using similar methods, but assumed that education does not change after 25 years of age and held education forecasts constant after 25 years of age within a given location-sex-specific birth cohort to prevent implausible within-cohort changes. These methods have been previously used in GBD forecasting papers on cause-specific mortality, and additional details can be found in these papers.,

21We forecasted the combined effect of the three GBD risk factors by calculating a risk factor scalar that describes the proportion of prevalence attributable to GBD risk factors. To calculate our risk factor scalar, we first estimated the population attributable fractions (PAFs) for each risk factor, using past and forecasted SEVs from 1990 to 2050 and relative risks describing the association between each risk factor and dementia, from GBD 2019. We then estimated the joint PAF for the three risk factors, taking into account potential mediation between the inter-related risk factors. Finally, we calculated scalars specific to each forecasted year, location, age group, and sex as:

22Additional details on the calculation of PAFs, risk factor mediation, and scalars can be found in the appendix (pp 21–22) or in previous GBD forecasting papers.,

23We then forecasted GBD risk-deleted prevalence, calculated as the total prevalence divided by the risk factor scalar. To select the additional risk factors to include in our forecasting model, we evaluated the association between each risk factor that was both quantified within the GBD framework and included in the 2020 Lancet Commission report (low physical activity, high systolic blood pressure, low education, alcohol use, and air pollution) and dementia prevalence (logit-transformed). We used sex-stratified models, controlling for 5-year age group and world region. We decided to include covariates for which the estimated association with dementia prevalence in both men and women was significant and in the same direction as the effect reported in the Lancet Commission report. On the basis of these criteria, we added education as a covariate in our forecasting model (appendix p 23).

24We used linear regression to forecast logit-transformed GBD risk-deleted prevalence, using sex-stratified models. We included covariates for 5-year age group (reference: 40–44 years), world region (reference: Andean Latin America; appendix p 61), and education (in years), as shown in the equation below depicting the linear regression models:

25Because our input data on prevalence and risk factors were themselves estimates, we ran 1000 models to incorporate uncertainty from our inputs. Each model used one draw from the distribution of each input to our forecasting model. Finally, to account for trends unexplained by our covariates, we fit a random walk model (autoregressive integrated moving average [ARIMA(0,1,0)]) to the residuals of our regression models and added forecasts of these residuals to our estimated forecasts.

26We calculated the total forecasted prevalence as the product of our forecasted prevalence not attributable to GBD risk factors and the forecasted risk factor scalar. To ensure continuity between past and future means and UIs, we intercept-shifted the forecasted draws of logit-transformed prevalence. We calculated the number of projected cases of dementia by multiplying prevalence forecasts with population forecasts, as calculated for the GBD population forecasting paper.

27To quantify the drivers of changes in the number of individuals with dementia, we estimated the relative contribution of population growth, population age structure, changes due to changes in GBD risk factor prevalence, and changes due to changes in education. For this decomposition analysis, we used methods developed by Das Gupta, which summarise the contribution of various factors to observed changes by algebraically isolating the standardised impact of each contributing multiplicative factor.

28Socio-demographic Index (SDI) is a composite indicator of total fertility rate, mean years of education in individuals older than 15 years, and lag-distributed income per capita. Countries were assigned to an SDI quintile on the basis of their estimated values on these indicators in 2019, and these designations were used to produce aggregated estimates of dementia prevalence by SDI quintile. We calculated age-standardised rates using the GBD world population standard. We propagated uncertainty through all components of the analysis by sampling 1000 draws at each stage of our estimation process. UIs were defined as the 25th and 975th values of the ordered draws. A difference in two quantities was defined as significant if the 95% UI of the difference did not include zero. For our analyses, we used Stata (versions 13.1 and 15.1), Python (versions 3.6.2 and 3.6.8), and R (versions 3.4.2 and 3.5.0).

29The study funders had no role in study design, data collection, data analysis, data interpretation, or the writing of the manuscript.

30We estimated a global percentage change in age-standardised both-sex dementia prevalence between 2019 and 2050 of 0·1% (95% UI −7·5 to 10·8), indicating that the global age-sex-specific prevalence of dementia is expected to remain stable. Dementia prevalence was higher in women than in men and increased with age, doubling about every 5 years until 85 years of age in both 2019 and 2050 (figure 1). In 2050, the prevalence of dementia in men globally was forecasted to be 0·5% (95% UI 0·4 to 0·7) among those aged 40–69 years, 6·5% (5·3 to 8·0) among those aged 70–84 years, and 23·5% (19·1 to 29·2) among those aged 85 years and older. In women, the global forecast was projected to be 0·6% (0·5 to 0·8) among those aged 40–69 years, 8·5% (7·0 to 10·4) among those aged 70–84 years, and 30·5% (25·0 to 36·8) among those aged 85 years and older (figure 1).

31In comparison, we estimated a much larger percentage change in all-age dementia prevalence between 2019 and 2050 (117% [95% UI 100–133]) due to population ageing, and an even larger percentage change in the all-age numbers of prevalent cases (166% [148–185]), due to a combination of population ageing and population growth (figure 2). In 2019, there were an estimated 57·4 (95% UI 50·4–65·1) million individuals living with dementia globally. We estimated that this number would increase to 83·2 (73·0–94·6) million individuals in 2030, 116·0 (100·7–132·1) million individuals in 2040, and 152·8 (130·8–175·9) million individuals living with dementia in 2050 (table). We estimated that there were more women with dementia than men with dementia globally in 2019 (female-to-male ratio of 1·69 [1·64–1·73]), and we expect this pattern to continue to 2050 (female-to-male ratio of 1·67 [1·52–1·85]).

32The percentage change between 2019 and 2050 in total number of dementia cases was highest for the low SDI quintile (330% [95% UI 295 to 362]) and lowest for the high-middle SDI quintile (113% [93 to 134]) and high SDI quintile (140% [117 to 160]). The percentage change in age-standardised rates between 2019 and 2050 was also highest for the low SDI quintile (8·2% [0·5 to 19·8]). The low-middle SDI quintile (1·3% [–6·9 to 12·4]) had the lowest percentage change in age-standardised rates. However, the observed differences between SDI quintiles in percentage change in age-standardised rates were not significant (table).

33When examining further patterns of percentage change in the number of projected dementia cases by country or world region, there was evidence of large variability in the projected increases. Although there were increases for every country, the smallest increases were projected in high-income Asia Pacific (53% [95% UI 41–67]) and western Europe (74% [58–90]), whereas the largest estimated increases were in north Africa and the Middle East (367% [329–403]) and eastern sub-Saharan Africa (357% [323–395]; figure 3; table).

34The decomposition of percentage change in projected dementia cases indicated a fairly small overall effect of changing prevalence, which included an increase in prevalence due to GBD risk factors for all regions except high-income Asia Pacific, and a decrease in prevalence due to the added effect of trends in education in this analysis. The increases attributable to GBD risk factors in most regions included the effects of increases in BMI and fasting plasma glucose, as well as less clear trends in smoking. Declines in dementia prevalence due to GBD risk factors in high-income Asia Pacific stem from expected declines in risk factor prevalence in this region. However, these effects were overwhelmed by changes due to population ageing and growth. Although at the global level population ageing and population growth contributed about equally to forecasted increases in dementia prevalence, we observed different patterns at the regional level. Whereas population growth is predicted to drive most of the increases in dementia prevalence in sub-Saharan Africa, population growth contributes to a much lesser extent in east Asia and western Europe. In central Europe, eastern Europe, and high-income Asia Pacific, anticipated declines in populations serve to partly counteract the effect of population ageing. Although population ageing is expected to lead to increases in projected dementia cases in every region, the largest projected increases due to population ageing are observed in east Asia, and north Africa and the Middle East (figure 4).

35Most of the projected increase in the numbers of people living with dementia can be attributed to population ageing and population growth, although the relative contribution of these two factors varied across GBD world regions. By contrast, we found that country-level projected changes in age-specific or age-standardised dementia prevalence attributable to changes over time in risk factor prevalence were fairly small. In most regions, increases in age-standardised dementia prevalence due to increases in the prevalence of GBD risk factors, including BMI, fasting plasma glucose, and smoking, were counterbalanced with decreases in age-standardised prevalence due to increases in average educational attainment.

36Our estimate for the global number of dementia cases in 2050 is similar to previous estimates from the World Alzheimer Report 2019 (152 million cases), but much higher than previous estimates of the forecasted global prevalence from work done by Brookmeyer and colleagues (106·8 million cases [95% UI 47·2–221·2]). The estimates from Brookmeyer and colleagues were only for Alzheimer's disease, which constitutes a subset of the total prevalence of dementia, and therefore it is expected that these estimates would be lower than the ones presented here. Due to the high prevalence of mixed pathologies among dementia cases and potential unknown differences in the distribution of underlying causes of dementia across geographical locations, the larger total dementia category is more relevant for the global study of dementia.

37Consistent with previous reports,, we estimated that there were more women with dementia than men with dementia in 2019, and our analysis indicated that this pattern would continue in 2050. These patterns exist despite the larger prevalence of vascular risks in men than in women, suggesting potentially strong counteracting mechanisms driving these inequities. Although the sex difference can be explained in part by higher life expectancy in women than in men, previous evidence also suggests potential sex differences in the biological mechanisms that underlie Alzheimer's disease.,

38We predicted that population ageing and population growth will drive enormous increases in the number of individuals affected by dementia both regionally and globally. The consequences of these increases are compounded by the high societal and monetary costs of the disease., Given the current absence of available effective disease-modifying treatments for dementia, immediate efforts to reduce these projected increases will need to target disease prevention through interventions for modifiable risk factors., Our analysis suggested that there would not be large changes in age-specific dementia prevalence due to changes in risk factor prevalence, based on the continuation of previous trends in the population-level exposure to educational attainment, BMI, fasting plasma glucose, and smoking. However, interventions that alter the expected trends in risk factor prevalence might reduce the expected future prevalence of dementia. Results from the 2020 update to the Lancet Commission on dementia prevention, intervention, and care suggest that up to 40% of dementia prevalence might be preventable through interventions targeting modifiable risk factors. This conclusion suggests that large changes in the exposure distributions of modifiable risk factors (ie, decreases in harmful risk factors and increases in protective factors) have the potential to considerably change our forecasted estimates and reduce the future burden of disease.

39The 2019 guidelines on dementia risk reduction from WHO further highlight potential interventions that might help to reduce the projected increases in dementia prevalence. However, these guidelines focus primarily on individual-level interventions. Our results indicate that there are differences in the estimated change in age-standardised prevalence rates across SDI quintiles, and suggest that structural inequalities must also be addressed to tackle the observed disparities.

40Our estimates can be used to plan for increases in the anticipated resources and support that will be needed by individuals with dementia and their caregivers in the future. Currently, many shortcomings exist in the identification, treatment, and care of individuals with dementia globally. Lower diagnosis rates in low-income settings than in high-income settings and gaps in the availability of resources for the management of neuropsychiatric symptoms of dementia will be important to address.,, Evidence suggests that hospitalisation rates are higher among those with dementia than among those without dementia, and health systems need to be prepared for the expected increases in service use that will accompany increases in the number of people with dementia. High rates of comorbidity with many other chronic diseases further complicate the care required for individuals with dementia and underscore the need to appropriately plan for the health-care needs of this population. Furthermore, although large gaps currently exist in the availability of quality end-of-life care for individuals with dementia, such care can positively affect both individuals with dementia and their caregivers., Therefore, efforts should be made to provide these services and scale up available resources to meet the needs of individuals with dementia in the future. The effect on the increasing numbers of caregivers that will be needed to support growing numbers of people with dementia should also be considered, and evidence-based interventions to support caregivers should be part of any comprehensive plan to address increases in dementia prevalence.

41Our study had some limitations. The methodology for forecasting prevalence due to GBD risk factors using PAFs relies on the assumptions of the PAF calculation—in particular, the assumption that risks examined are causally related to dementia. To the extent that the estimated relative risks reflect non-causal increases or decreases in risk, this assumption might be violated. Our estimates of dementia prevalence from the GBD study are based on sparse and heterogeneous data because of the absence of standardised methodologies for measuring dementia in population-based studies. When no data existed, we relied on the Bayesian framework of the DisMod-MR 2.1 and assumed that prevalence could be explained using regional estimates of prevalence adjusted by covariate values. Although there is no available gold standard to which we can judge the validity of our estimates, our global estimates are very similar to those from other efforts, such as the World Alzheimer Report. Furthermore, some of the geographical heterogeneity observed could be due to methodological differences in prevalence studies, even though we attempted to adjust for studies using different methodologies and different definitions of dementia. Due to the heterogeneity in prevalence data over time, we did not allow observed time trends to affect our forecasted time trends in the future, because this could have induced spurious trends caused by changes in diagnostic procedures over time. Instead, our future time trends are based solely on expected trends in risk factor prevalence, as well as trends in population ageing and growth. Further investments in high-quality epidemiological studies are required to better estimate trends over time in dementia incidence and prevalence as well as geographical variation in dementia prevalence. This observed heterogeneity in prevalence data, in combination with high correlations between GBD risk factors and other dementia risk factors, could have had a role in our ability to detect a significant association between risk factors of dementia that are not currently included in the GBD study and GBD risk-deleted dementia prevalence by country. This limited the number of risk factors we could include as predictors of future dementia prevalence as well as the risk factors we could include in the decomposition analysis showing the effect of risks on future prevalence. The number of risk factors we tested for inclusion was additionally limited to include only those risk factors that were included in the 2020 Lancet Commission report and are part of the GBD comparative risk assessment framework., Despite this limitation, we were able to include risks for which large changes in the prevalence of exposure are expected, including BMI, educational attainment, and fasting plasma glucose.,, These risks are of most importance to future forecasts, which will only be affected by the exclusion of risk factors if the future exposure to a given risk is expected to change. However, future efforts should seek to incorporate more risks within GBD, which would allow us to include additional risk factors without relying on the estimation of the ecological association between risk exposure prevalence and dementia prevalence.

42We also assumed that the maximal educational attainment was 18 years of education and that educational attainment was constant after the age of 25 years, as higher levels of education are probably not consequential for most health outcomes quantified in the GBD project. However, continued education throughout the life course might convey a greater protective benefit for dementia. Furthermore, our exposure estimates for high BMI are based on the distribution of BMI in a population at a specific timepoint, but the evidence for BMI as a risk factor suggests that midlife BMI is most important. Therefore, future efforts should consider lagging BMI exposure to capture the risk–outcome association in forecasts more accurately. Additionally, this study focused on the overall prevalence of dementia because of a scarcity of globally relevant data on the decomposition of the broader dementia category into clinical subtypes. It is possible that clinical subtypes, such as vascular dementia, might have different underlying associations with dementia risk factors, which could affect our results. However, evidence suggesting a large overlap of neuropathologies across clinical subtypes calls into question the precision or even utility of clinical subtype designations., Our study also focused on the estimation of dementia prevalence, as the estimation of the numbers of individuals with dementia is of direct relevance for public health planning efforts. However, the estimation of future dementia incidence is also useful and should be the topic of future investigations.

43Despite these limitations, this study improves on previous global forecasts of dementia in several important ways. First, drawing on the strength of the country-level estimates from GBD 2019, we were able to produce estimates of projected prevalence by country, world region, and SDI quintile, whereas previous studies have been restricted to global and regional estimates. Second, we used country-level forecasts of four different risk factors that have been shown to be associated with dementia, and the observed association between these risks and dementia prevalence from 1990 to 2019, to incorporate changes in prevalence that would be due to changes in these risk factors. Third, we were able to quantify through decomposition analysis the relative contribution of changes in GBD risk factors, changes in education, changes in population growth, and changes in population ageing.

44The country-level specificity of our estimates will allow policy makers and decision makers to understand the expected increases in the number of individuals with dementia and the drivers of these increases in a given geographical setting. This information might be helpful for public health planning efforts, particularly as they relate to scaling up the availability of resources required to meet the needs of individuals with dementia and their caregivers. Furthermore, the projected increases in the number of people with dementia, due largely to population growth and population ageing, underscore the crucial need for research focused on the discovery of disease-modifying treatments, effective low-cost interventions, and novel modifiable risk factors for the prevention or delay of disease onset.

45Correspondence to: Emma Nichols, Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98105, USA eln1@uw.edu

46Correspondence to: Emma Nichols, Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98105, USA eln1@uw.edu

47Citations for the data used in the study can be accessed from the Global Health Data Exchange (http://internal-ghdx.healthdata.org/). Access to the data is also provided as a data use agreements permit.

48Y Bejot reports honoraria for lectures from Boehringer-Ingelheim, Bristol Myers Squibb, Pfizer, Servier, Medtronic, and Amgen, outside the submitted work. C Brayne reports support for the present manuscript from the Economic and Social Research Council (ESRC), Alzheimer's Research UK, National Institute for Health Research (NIHR), Medical Research Council (MRC), Alzheimer's Society, East Anglia Regional Health Authority Public Health and Operational Research Advisory Council, Regional Health Authority (as research grants from 1990 to date for the Cognitive Function and Ageing Studies), paid to their institution; honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from the National Institute on Aging (NIA) Health and Retirement Study (HRS) Data Monitoring Committee and AXA Research Fund Scientific Board, paid to their institution, and from the Department of Biotechnology and Wellcome Trust India Alliance Fellowship Selection Committee, as a personal payment; support for attending meetings and travel from NIA HRS Data Monitoring Committee, Department of Biotechnology and Wellcome Trust India Alliance Fellowship Selection Committee, Canadian Longitudinal Study on Aging Scientific Advisory Board, Alzheimer's Society Research Strategy Council, BRAIN & HEADING International Oversight Committee, The Irish Longitudinal Study on Aging (TILDA) Scientific Advisory Board, ATHLOS Advisory Board, Barcelona Brain Health Initiative Scientific Advisory Board, DZNE International Scientific Review Panel (Humboldt), Faculty of Public Health Academic & Research Committee, and Faculty of Public Health Board; participation on a Data Safety Monitoring Board or Advisory Board with NIA HRS Data Monitoring Committee, AXA Research Fund Scientific Board, Department of Biotechnology and Wellcome Trust India Alliance Fellows, Canadian Longitudinal Study on Aging Scientific Advisory Board, Alzheimer's Society Research Strategy Council, BRAIN & HEADING International Oversight Committee, TILDA Scientific Advisory Board, Chinese University of Hong Kong Project Advisory Board, University of Sheffield Health Lifespan Institute Advisory Board, ATHLOS Advisory Board, Barcelona Brain Health Initiative Scientific Advisory Board, DZNE International Scientific Review Panel (Humboldt), Scientific Advisory Board for UKPRP Air Pollution and Cognitive Health Consortium, and InSPIRE; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid, with Faculty of Public Health Academic & Research Committee as Chair, Faculty of Public Health Board as a trustee, Public Health England–University of Cambridge Academic Liaison Committee meeting as Chair, and East of England Public Health England Research and Evaluation Hub as co-Chair; all outside the submitted work. I Filip reports payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Avicenna Medical and Clinical Research Institute in the form of financial support, outside the submitted work. B J Hall reports consulting fees from WHO, and holds a US S&P index fund and a US Bond Index fund, all outside the submitted work. C Herteliu reports grants from the Romanian National Authority for Scientific Research and Innovation (CNDS-UEFISCDI, project number PN-III-P4-ID-PCCF-2016-0084), outside the submitted work. A Kandel reports grants from the University at Buffalo Clinical and Translational Institute. M Kivimäki reports support for the present manuscript from the MRC (S011676) and the Wellcome Trust (221854/Z/20/Z) as a grant paid to their institution. S Lorkowski reports grants or contracts from Akcea Therapeutics as payments made to their institution; consulting fees from Danone, Swedish Orphan Biovitrum (SOBI), and Upfield; payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Akcea Therapeutics, AMARIN, Amedes Holding, Amgen, Berlin-Chemie, Boehringer-Ingelheim Pharma, Daiichi Sankyo Deutschland, Danone, Hubert Burda Media Holding, Lilly Deutschland, Novo Nordisk Pharma, Roche Pharma, Sanofi-Aventis, and SYNLAB Holding and SYNLAB Akademie as personal payments; support for attending meetings and travel from Amgen as personal payments; participation on a Data Safety Monitoring Board or Advisory Board with Akcea Therapeutics, Amgen, Daiichi Sankyo, and Sanofi-Aventis as personal payments; all outside the submitted work. J Massano reports consulting fees from Roche, Biogen, Bial, and AbbVie; payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Bial, GE Healthcare, Boston Scientific, and Merck Sharp & Dohme; support for attending meetings and travel from Bial and Roche; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid, with the Portuguese Brain Aging and Dementia Study Group as President; all outside the submitted work. C D Pond reports payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Dementia Training Australia; leadership or fiduciary role in other board, society, committee, or advocacy group, unpaid, as an advisor for the Primary Health Network. A Radfar reports payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Avicenna Medical and Clinical Research Institute. J A Singh reports consulting fees from Crealta–Horizon, Medisys, Fidia, Two Labs, Adept Field Solutions, Clinical Care options, Clearview healthcare partners, Putnam associates, Focus forward, Navigant consulting, Spherix, MedIQ, UBM, Trio Health, Medscape, WebMD, Practice Point communications, the NIH, and the American College of Rheumatology; payment or honoraria for lectures, presentations, speakers' bureaus, manuscript writing, or educational events from Simply Speaking; support for attending meetings and travel from OMERACT, an international organisation that develops measures for clinical trials and receives arm's length funding from 12 pharmaceutical companies; participation on a Data Safety Monitoring Board or Advisory Board as a member of the Food and Drug Administration Arthritis Advisory Committee; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid, with OMERACT as a member of the steering committee, with the Veterans Affairs Rheumatology Field Advisory Committee as a member, and with the UAB Cochrane Musculoskeletal Group Satellite Center on Network Meta-analysis as a director and editor; stock or stock options in TPT Global Tech, Vaxart pharmaceuticals, Charlotte's Web Holdings, and previously owned stock options in Amarin, Viking, and Moderna pharmaceuticals; all outside the submitted work. D J Stein reports personal fees from Lundbeck, Takeda, Johnson & Johnson, and Servier, outside the submitted work. A Wimo reports support for the present manuscript from WHO as payment to their institution and from the Swedish Government (SNAC project) paid to their county council; grants or contracts from Merck Sharp & Dohme (research grant, EU-project IMI2: MOPEAD, EU-project H2020; PRODEMOS, EU-project JPND: MindAD), paid to their institution; royalties and licenses with an RUD instrument as a partial license holder; support for attending meetings and travel to Geneva from WHO, and to Seattle, WA, USA, from the Institute of Health Metrics and Evaluation (IHME); participation on a Data Safety Monitoring Board or Advisory Board with Biogen, Elsai, and IHME; all outside the submitted work.