Unraveling disparities in county‐level dementia diagnosis prevalence across the United States

Abstract

INTRODUCTION

We aimed to identify demographic, socioeconomic, and environmental factors contributing to county‐level variation in dementia diagnosis prevalence across the United States.

METHODS

Using 2020 Dementia Data Hub data covering 61.5 million Medicare beneficiaries, we modeled the top tertile of dementia diagnosis prevalence across 3076 counties. Forty‐one county‐level variables were evaluated using logistic regression and region‐specific models.

RESULTS

Top‐tertile counties averaged 1986 dementia cases per 100,000 residents; 43.8% were in the South. The main model included seven predictors: higher diabetes prevalence, uninsured rate, fast‐food access, White race prevalence, smaller household size, smoking rate, and elevation (area under the curve [AUC]: 0.84; 95% confidence interval: 0.82 to 0.85). Region‐specific models improved accuracy (AUCs 0.83 to 0.89).

DISCUSSION

Dementia diagnosis prevalence varies widely across the United States and can be predicted with high accuracy using a small set of regionally adaptable variables. Region‐specific modeling may help policymakers identify high‐burden communities, tailor prevention strategies, and monitor the impact of targeted interventions over time.

Highlights

• Six models tested predictors of county‐level dementia diagnosis prevalence.

• Backward selection was the main model: seven variables, high accuracy (AUC = 0.84).

• Higher diabetes, uninsured rates, White race prevalence, and fast food linked to more dementia diagnoses.

• Larger households, higher elevation, and more smokers linked to less dementia diagnoses.

• Some predictors (e.g., elevation) likely act as proxies; collinearity limits causality.

1. BACKGROUND

Dementia is a leading cause of disability and dependency among older adults. The prevalence is rising as our global population ages and is anticipated to triple from 50 million in 2025 to 150 million by 2050, ¹ which will create significant challenges for public health systems worldwide. ² , ³ , ⁴ , ⁵ , ⁶ , ⁷

Up to 40% of dementia risk is modifiable and could be prevented through risk factor management, ¹ , ⁸ , ⁹ which highlights the importance of understanding the drivers of dementia at a granular level. While prior research provided insights into dementia prevalence within specific populations or nationally using survey data, ¹⁰ , ¹¹ , ¹² less is known about how the burden of dementia diagnoses vary at smaller geographical units across the United States. ¹³ A recent study applied a dementia prevalence estimate from a single metropolitan area to all US counties but did not directly measure prevalence. ¹⁴ In addition, prior research did not fully consider the complex interplay of demographic, socioeconomic, medical, and environmental factors that contribute to county‐level dementia diagnosis prevalence. The degree to which predictive models incorporating these contributors might perform differently by US region also remains underexplored.

To address these gaps, we leveraged the Dementia Data Hub (DDH), a newly launched resource that provides the first publicly accessible county‐level estimates of diagnosed dementia in the United States. The DDH integrates claims data from over 61 million Medicare beneficiaries, including both fee‐for‐service (FFS) and Medicare Advantage enrollees, across diverse geographic and democratic contexts. By harmonizing DDH data with complementary governmental and non‐governmental sources, we systematically evaluated the contributions of 41 county‐level factors to dementia diagnosis prevalence in 2020.

Our study attempted to determine how variations in risk factor prevalence contributed to county‐level disparities in dementia diagnosis prevalence. Answering this question could have substantial implications for targeted public health interventions aimed at reducing incident dementia. The findings are not intended to imply mechanistic or causal relationships. By focusing on parsimonious predictive models, these analyses may provide actionable targets for policymakers and health systems.

2. METHODS

2.1. Dataset

Data were sourced from the DDH, ¹⁵ a partnership between NORC at the University of Chicago and the Milken Institute School of Public Health. The dataset encompasses 61.5 million Medicare beneficiaries from both the FFS and Medicare Advantage (MA) programs. The beneficiaries included in the dataset had continuous Medicare enrollment between 2017 and 2019. All DDH data used in this study are from 2020, and dementia diagnoses were determined using a previously published algorithm with three levels of diagnostic certainty: highly likely, likely, or possible dementia. ¹⁶ The DDH then aggregates measures of dementia incidence and prevalence at state and county levels.

2.2. Outcome

The study outcome is county‐level prevalence of highly likely dementia, which DDH defined as beneficiaries with at least two claims or encounters on different dates containing International Classification of Diseases, 10th Revision, Clinical Modification codes from the most frequently used and validated tiers (e.g., Alzheimer's disease or vascular dementia codes) (Table S1). ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² Beneficiaries were further classified using prescription drug data from Medicare Part D for dementia‐targeting medications, because such drugs reinforced the likelihood of a diagnosis. The approach applied a 3‐year lookback period (2017 to 2019) to ensure comprehensive capture of medical history while minimizing false positives. This hierarchical classification aimed to reflect a true dementia diagnosis rather than ambiguous or unrelated symptoms. To reduce confounding, prevalence rates were age‐standardized to the general Medicare population. Nonetheless, the dementia category includes a broad spectrum of dementia types or severity and does not include cases diagnosed in individuals who did not have Medicare or undiagnosed dementia cases.

RESEARCH IN CONTEXT

1. Systematic review: We reviewed PubMed to identify studies on geographic patterns in dementia prevalence and associated county‐level risk factors. Although prior work linked dementia disparities to demographic and socioeconomic variables, few studies directly measured county‐level dementia diagnosis prevalence or evaluated the predictive value of county‐level risk factors.

2. Interpretation: Using county‐level demographic, medical, socioeconomic, behavioral, and environmental data, we developed models that accurately predicted counties with a high dementia burden. The main model included diabetes prevalence, number of household members, uninsured rate, fast‐food access, smoking rate, White race prevalence, and elevation. The high collinearity of predictor variables suggest they may serve as proxies for more complex contextual influences. Region‐specific models improved accuracy, highlighting the value of localized approaches for dementia surveillance and prevention.

3. Future directions: Future research should disentangle mechanistic pathways underlying these proxy predictors and integrate local context to improve predictive modeling accuracy and practical utility.

2.3. Exposures

While the outcome relied solely on DDH data, we leveraged 10 publicly available data sources to define county‐level exposures, which are detailed in Table 1. We used data from 2020 or as close to that year as possible and then harmonized those data with the DDH outcome.

TABLE 1.

County‐level variables, year of data, source, and other relevant information.

Variable	Description	Year	Source	Notes
Demographic exposures
Age	Median age	2020	Decennial Census	Self‐reported
White	Percentage of population self‐identifying as Non‐Hispanic White	2020	Decennial Census	Excludes Hispanics
Black	Percentage of population self‐identifying as Non‐Hispanic Black	2020	Decennial Census	Excludes Hispanics
Hispanic	Percentage of population self‐identifying as Hispanic	2020	Decennial Census	Excludes another race/ethnicity
Education	Percentage of population with less than high school education (ages ≥25 years)	2020	American Community Survey	AHRQ Social Determinants of Health Database
Socioeconomic exposures
Income	Median household income	2020	Decennial Census	Small Area Income and Poverty Estimates
Poverty	Percentage of population below federal poverty line	2020	Decennial Census	Small Area Income and Poverty Estimates
Unemployment	Unemployment rate	2020	Bureau of Labor Statistics	For population ≥16 years
Gini	Gini index of income inequality	2020	American Community Survey	AHRQ Social Determinants of Health Database
SVI	SVI	2020	CDC	CDC/ATSDR SVI
SDI	Social Deprivation Index	2019	Robert Graham Center	Relies on similar data as SVI
Digital divide	Percentage of households without Internet access	2016 to 2020	American Community Survey	CDC/ATSDR SVI
Crowded housing	Percentage of housing units with more than one occupant per room	2020	American Community Survey	AHRQ Social Determinants of Health Database
Household size	Average number of people in occupied housing units	2020	American Community Survey	AHRQ Social Determinants of Health Database
Public assistance	Percentage of households with public assistance income or food stamps/SNAP	2020	American Community Survey	AHRQ Social Determinants of Health Database
Medical Exposures
CHD	Crude prevalence of adults with coronary heart disease	2020	Behavioral Risk Factor Surveillance System	CDC PLACES
Diabetes	Crude prevalence of adults with diabetes	2020	Behavioral Risk Factor Surveillance System	CDC PLACES
Stroke	Crude prevalence of adults with stroke	2020	Behavioral Risk Factor Surveillance System	CDC PLACES
Hypertension	Crude prevalence of adults with hypertension	2019	Behavioral Risk Factor Surveillance System	CDC PLACES
Hearing Loss	Crude prevalence of adults with hearing disability	2021	Behavioral Risk Factor Surveillance System	CDC PLACES
Poor sleep	Crude prevalence of adults who sleep < 7 h per night	2020	Behavioral Risk Factor Surveillance System	CDC PLACES
Disabled	Percent of population with a disability	2020	American Community Survey	AHRQ Social Determinants of Health Database
Physical Inactivity	Crude prevalence of adults with no leisure‐time physical activity	2020	Behavioral Risk Factor Surveillance System	CDC PLACES
Physicians	Percent of population that are active physicians	2020	American Medical Association Physician Master File	HRSA Area Health Resources Files
Neurologists	Percent of population that are active neurologists	2020	American Medical Association Physician Master File	HRSA Area Health Resources Files
Primary Care Visit	Percent of Medicare Enrollees Having Annual Visit to a Primary Care Clinician	2019	100% sample of Medicare fee‐for‐service	Dartmouth Atlas Data
Hospital Beds	Ratio of total hospital beds to population	2020	American Hospital Association Survey Database	HRSA Area Health Resources Files
Hospital Admissions	Ratio of all hospital inpatient admissions to population	2020	American Hospital Association Survey Database	HRSA Area Health Resources Files
Population Density	Number of inhabitants per square mile	2020	Decennial Census	Does not include migratory populations
Medicare Reimbursement	Per capita Medicare reimbursements	2019	100% sample of Medicare fee‐for‐service	Dartmouth Atlas Data
Uninsured	Crude prevalence of adults without insurance aged 18 to 64 years	2020	Behavioral Risk Factor Surveillance System	CDC PLACES
Environmental Exposures
Air Quality	Average PM2.5 concentrations in µg/m3	2018	National Environmental Public Health Tracking Network	CDC
Near highway	Percentage of population living within 150 m of a highway	2020	Federal Highway Administration Functional Class System	Environmental Public Health Tracking Network
Walk/Bike	Percentage of workers >16 years of age that used active transportation	2020	American Community Survey	Specific to the commute to work
Fast food	Fast‐food restaurants/1000 population	2016	Food Environment Atlas	United States Department of Agriculture
Famers markets	Farmers’ Markets/1000 population	2018	Food Environment Atlas	United States Department of Agriculture
Food dessert	Percentage of population with low access to store that sells food	2015	Food Environment Atlas	United States Department of Agriculture
Fitness Facilities	Recreation and fitness facilities/1000 population	2016	Food Environment Atlas	United States Department of Agriculture
RUCC	Rural–Urban Continuum Codes	2023	Economic Research Service	United States Department of Agriculture
Temperature	Average annual temperature	2020	National Oceanic and Atmospheric Administration	US Department of Commerce
Elevation	Average elevation above sea level	2014	Environmental Systems Research Institute	US Geological Survey

Abbreviations: AHRQ, Agency for Healthcare Research and Quality; ATSDR, Agency for Toxic Substances and Disease Registry; CDC, Centers for Disease Control and Prevention; HRSA, Health Resources and Services Administration; SNAP, Supplemental Nutrition Assistance Program; SVI, Social Vulnerability Index; USDA, US Department of Agriculture.

Demographic variables included median age, race/ethnicity distribution (non‐Hispanic White, non‐Hispanic Black, Hispanic), and education level (percentage of the population with less than a high school education), obtained from the 2020 Decennial Census and the American Community Survey. ²³ Socioeconomic variables included median household income, poverty rate, unemployment rate, percentage of households receiving public assistance, Gini index of income inequality, ²⁴ Social Vulnerability Index (SVI), ²⁵ Social Deprivation Index (SDI), ²⁶ indicators of the digital divide (households without Internet access), average household size, and crowded housing metrics, sourced from the Census, Agency for Healthcare Research and Quality (AHRQ) databases, and the Centers for Disease Control and Prevention (CDC). ²⁷

Medical exposures included prevalence of diabetes, hypertension, coronary heart disease (CHD), stroke, hearing loss, disability, rates of uninsured adults, physical inactivity, and poor sleep. These data were extracted from the 2020 Behavioral Risk Factor Surveillance System (BRFSS), ²⁸ CDC PLACES, ²⁹ and AHRQ datasets. ³⁰ Additional medical exposures included physician density, neurologist density, ratio of hospital beds to population, inpatient admission rates, average Medicare reimbursement per beneficiary, and the percentage of Medicare enrollees having an annual visit to a primary care clinician. These data were derived from the American Medical Association Physician Master File, Health Resources and Services Administration (HRSA), Dartmouth Atlas, and American Hospital Association databases. ³¹

Environmental exposures included air quality (PM2.5 concentration), proximity to highways, average annual temperature, average elevation, transportation behaviors (percentage of workers using active transportation), population density, and the rural‐urban continuum code. Additional community resource variables encompassed fast‐food density, farmers’ market availability, food deserts, and recreation and fitness facility availability. Data sources for these variables included the National Environmental Public Health Tracking Network, ³² Federal Highway Administration, ³³ United States Geological Survey, ³⁴ National Oceanic and Atmospheric Administration, ³⁵ and US Department of Agriculture Food Environment Atlas. ³⁶

2.4. Statistical analysis

Attempts to use linear regression resulted in an inability to achieve acceptable model fit (Figure S1). Given the right skew of the dependent variable (Figure S2), we modeled the top tertile of dementia diagnosis prevalence as the primary outcome. The top tertile of dementia diagnosis prevalence is a clinically meaningful outcome that is straightforward to interpret and can be used to identify outlier counties. Independent variables were standardized into z‐scores to ensure consistent scaling and comparability. To help visualize the dataset, we generated a correlation matrix where each cell contained the absolute value of the Pearson correlation coefficient between all variables in the dataset. We also generated a second graphical representation that visualized the strength and direction (positive or negative) of the correlations using color intensity or size.

To identify predictor variables, we used six approaches to model building: a priori variable selection, backward selection, partial correlation with stepwise regression, ridge/least absolute shrinkage and selection operator (LASSO) regularization, cross‐validated logistic LASSO, and principal component analysis (PCA). A priori variable selection involved including variables based on theoretical and clinical relevance. Backward selection (p < 1×10⁻⁹ selected for stability) systematically removed variables with the least statistical significance until the best‐fitting model was identified. Partial correlation with stepwise regression combined statistical selection with correlation analysis to refine variable inclusion. Ridge and LASSO regularization methods helped address multicollinearity and improve model generalizability by penalizing less relevant predictors. Cross‐validated logistic LASSO further optimized variable selection by minimizing overfitting through repeated subsampling and validation. Finally, PCA reduced dimensionality by identifying orthogonal components representing shared variance among predictors. Each approach provided complementary insights to robustly identify key predictors while mitigating bias and overfitting.

To evaluate the performance and utility of different variable selection approaches, we compared the six models based on metrics that assessed predictive accuracy, model fit, parsimony, and multicollinearity. The area under the curve (AUC) was used to measure model discrimination, reflecting the ability of each model to differentiate between outcomes. Model fit was assessed using the Bayesian Information Criterion (BIC) and Akaike information criterion (AIC), where lower values indicated better fit while penalizing for complexity. ³⁷ Accuracy, expressed as the percentage of correctly classified outcomes, provided an additional measure of predictive performance. McFadden's pseudo‐R ² was used to indicate the explanatory power of the models, ³⁸ reflecting how well the variables accounted for the variability in the outcome. The Hosmer–Lemeshow statistic assessed goodness of fit, while the mean variance inflation factor (VIF) quantified multicollinearity among predictors. ³⁹ , ⁴⁰ Finally, the number and type of variables included in each model highlighted differences in parsimony and predictor selection across approaches. Together, these metrics provided a comprehensive comparison of model performance across the six selection strategies.

We used the backward selection model for the main model because it provided a strong balance between predictive performance, model fit, and interpretability. It also demonstrated excellent discrimination and a good fit to the data, while maintaining parsimony compared to more complex models. The backward selection approach yielded a model with seven variables: household members, uninsured rate, diabetes prevalence, fast‐food access, smoking prevalence, White prevalence, and elevation. By selecting a limited set of predictors, the backward selection model offers a practical and interpretable framework for understanding the relationships between these variables and the outcome.

We performed a sensitivity analysis using two additional modeling approaches. The first alternative model employed a mixed‐effects logistic regression framework, with Census Region (Midwest, Northeast, South, and West) included as a random effect to account for potential geographic clustering. The second model extended this approach by incorporating population weighting within each region to further adjust for differences in population size across regions. These models allowed us to assess the influence of regional and population‐level factors on the estimated associations while comparing results with the primary logistic regression model.

We also assessed the performance of the main model separately within each Census Region to evaluate its predictive accuracy and consistency across geographic areas. For each region, we calculated the AUC along with 95% confidence intervals (CIs) to measure model discrimination. We then refined the main model by defining the included variables with backward selection (p < 0.01) for each Census Region to account for regional differences in socioeconomic, demographic, and health factors. The tailored models for each region were evaluated using AUC and corresponding 95% CIs to compare their performance. This approach allowed us to assess the impact of region‐specific variables on model performance while highlighting geographic variability in key predictors.

To further evaluate the performance of both the main model and the regionally derived model, we assessed classification accuracy. For the main model, we applied the primary analysis set of predictors across all regions without incorporating any region‐specific adjustments. Separately, for the regionally derived model, we used the predictors specific to each Census Region. We calculated region‐specific probabilities to classify outcomes (predicted class) and compared them to observed outcomes. We then generated a binary classification (predicted class) based on a threshold of 0.5. To determine whether classification accuracy differed significantly between regions, we stratified the results by Census Region.

As another sensitivity analysis, we used quantile regression to capture relationships across the outcome distribution. Variables were selected using stepwise backward selection, and results for the 25th, 50th, and 75th quantiles, including coefficients and 95% CIs, are presented. This method addressed issues with non‐normality and heteroscedasticity, providing a more robust framework for comparison to the main model analysis.

As a final sensitivity analysis, we examined county adjacency by comparing the counties in the top tertile of dementia diagnosis from our primary analysis to any adjacent counties with lower dementia diagnosis. County adjacency relationships were obtained from the US Census Bureau's County Adjacency File, ⁴¹ which identifies counties sharing common boundaries or vertices. We excluded self‐adjacencies and limited the comparison group to neighboring counties with lower dementia diagnosis. Adjacent counties were deduplicated to ensure each unique county appeared only once in the analysis. Statistical comparisons were performed using two‐sample t tests for the individual covariates of the main model: household members, uninsured rate, diabetes prevalence, fast‐food access, smoking prevalence, White prevalence, and elevation.

The main model has data for 3076 (97.9%) of the 3143 counties in the United States. Some counties have suppressed data given small cell counts from low total population. Given the limited number of counties with missing data (n = 67), we did not impute data. All analysis was performed in Stata 18.0 (StataCorp, College Station, TX, USA), except for the correlation matrix figure, which was created in Anaconda (Anaconda, Austin, TX, USA). A two‐sided p value < 0.05 was considered statistically significant.

3. RESULTS

Of the 3076 included counties, 991 counties (32.2%) ranked in the top tertile of dementia diagnosis prevalence with 1986 ± 430 cases per 100,000 residents compared to 1244 ± 1232 in the lowest/middle tertile. A violin plot of the primary outcome is presented in Figure S1. The highest burden of dementia was in the Southern Census Region, with 43.8% of Southern counties falling in the top tertile of diagnosis prevalence, compared to 27.8% of Midwestern counties, 23.0% of Northeastern counties, and 9.4% of Western counties (p < 0.001 for difference, Figure 1).

FIGURE 1.

Primary outcome (top tertile of dementia diagnosis prevalence) shown for US counties on map of lower 48 states.

The mean values of the standardized study exposures are seen after stratification by the outcome in Table 2. In order of magnitude, the top 10 largest differences in strata of the primary outcome were seen for the exposure variables of hypertension prevalence, CHD prevalence, stroke prevalence, diabetes prevalence, median age, household income, disabled prevalence, hearing loss, inactivity, and poverty prevalence. Figure 2 shows all the pairwise correlations in the dataset. Overall, there are 128 unique correlations with a correlation coefficient > 0.5 in this dataset, highlighting the extensive collinearity. To further illustrate this point, Figure S3 also shows the correlations, but using size and color to show the strength and directionality of the correlation.

TABLE 2.

Mean values for exposure variables after stratification by primary outcome.

Variable	Lower tertile dementia diagnosis	Top tertile dementia diagnosis	Delta	P value
Unemployment rate	0.00	0.03	0.03	0.411
Food desert	−0.06	−0.03	0.03	0.357
Farmers market access	−0.02	0.03	0.05	0.191
Air quality	0.04	−0.02	0.06	0.128
Near highway	0.04	−0.02	0.06	0.118
Neurologist density	0.04	−0.07	0.10	0.007
Hospital admissions	−0.02	0.07	0.10	0.013
Fast‐food access	− 0.02	0.09	0.10	0.004
White prevalence	0.05	0.15	0.11	0.002
Population density	0.03	−0.09	0.12	0.003
Walk/Bike for transport	−0.02	−0.14	0.12	<0.001
Medicare reimbursement	−0.02	0.10	0.13	<0.001
Physician density	0.04	−0.10	0.14	<0.001
Hospital beds	−0.04	0.13	0.18	<0.001
Fitness facilities	0.08	−0.14	0.22	<0.001
Crowded housing	0.05	−0.19	0.24	<0.001
Hispanic prevalence	−0.05	−0.19	0.24	<0.001
Education	−0.13	0.13	0.26	<0.001
Uninsured rate	− 0.08	0.17	0.26	<0.001
Social Vulnerability Index	−0.08	0.19	0.27	<0.001
Public assistance	−0.19	0.10	0.28	<0.001
Black prevalence	−0.07	0.21	0.28	<0.001
Average elevation	0.08	− 0.22	0.30	<0.001
Poor sleep	−0.09	0.22	0.31	<0.001
Social Deprivation Index	−0.09	0.21	0.31	<0.001
Primary care visit	−0.09	0.23	0.32	<0.001
Gini coefficient	−0.14	0.21	0.34	<0.001
Smoking prevalence	− 0.13	0.27	0.40	<0.001
Average temperature	−0.13	0.28	0.41	<0.001
Rural–urban classification	−0.14	0.29	0.43	<0.001
Household members	0.12	−0.35	0.47	<0.001
No Internet	−0.17	0.32	0.48	<0.001
Poverty prevalence	−0.16	0.33	0.49	<0.001
Inactivity	−0.18	0.39	0.57	<0.001
Hearing loss	−0.22	0.40	0.62	<0.001
Disabled prevalence	−0.24	0.42	0.65	<0.001
Household income	0.22	−0.45	0.67	<0.001
Median age	−0.26	0.45	0.71	<0.001
Diabetes prevalence	− 0.26	0.52	0.78	<0.001
Stroke prevalence	−0.27	0.55	0.82	<0.001
Coronary heart disease prevalence	−0.30	0.59	0.88	<0.001
Hypertension prevalence	−0.29	0.59	0.88	<0.001

Note: Exposure variables included in final model are bolded and italicized.

FIGURE 2.

Pairwise correlations of all variables in analysis, shown by order of correlation strength with primary outcome of top tertile of dementia diagnosis prevalence, without direction of correlation.

The six approaches to model building, detailed in Table 3, demonstrate the performance and trade‐offs of various methods in predicting the outcome. All models achieved good discrimination (AUC ≥ 0.8), with the cross‐validated logistic LASSO achieving the highest AUC of 0.85 and an accuracy of 78.2% but included 18 predictor variables. The backward selection model offered comparable discrimination and accuracy (AUC = 0.84, accuracy = 77.5%), while requiring only seven predictor variables.

TABLE 3.

Comparison of the six different model building approaches.

Model	AUC	BIC	AIC	Acc.	Pseudo R 2	Hosmer–Lemeshow	Mean VIF	No. of Var	Variables
A priori model	0.80	3216	3078	74.9%	0.21	0.02	2.39	7	Median Age, Black Prevalence, Education, Poverty, Uninsured Rate, Unemployment rate, HTN
Backward selection	0.84	2911	2863	77.5%	0.26	0.43	2.01	7	Household Members, Uninsured Rate, Diabetes, Fast‐Food Access, Smoking, White Prevalence, Elevation
Partial correlation with stepwise	0.84	2830	2740	78.4%	0.28	0.03	3.28	14	Median Age, Income, Physician Density, Uninsured Rate, Black Pop, Near Highway, Fast‐Food Access, Smoking, Diabetes, HTN, Hearing Loss, Elevation, Medicare Reimbursement, Primary Care Visit
Ridge/LASSO regularization	0.84	2867	2783	78.2%	0.28	0.38	2.78	13	Median Age, Household Members, Crowded Housing, Income, Uninsured Rate, CHD, HTN, Hospital Beds, Fast‐Food Access, Elevation, Temperature, Medicare Reimbursement, Primary Care Visit
Cross‐validated logistic LASSO	0.85	2878	2776	78.2%	0.28	<0.001	4.01	18	Median Age, Household Members, Income, Uninsured Rate, Physician Density, Hospital Admissions, Hospital Beds, Crowded Housing, Fast‐Food Access, Near Highway, CHD, Diabetes, HTN, Smoking, Temperature, Elevation, Medicare Reimbursement, Primary Care Visit
Principal component analysis	0.83	2806	2782	77.3%	0.25	<0.001	1.00	3	PC1, PC2, PC3

Abbreviations: Acc., accuracy; AIC, Akaike information criterion; AUC, area under the curve; BIC, Bayesian Information Criterion; CHD, coronary heart disease; HTN, hypertension; LASSO, least absolute shrinkage and selection operator; PC, principal component; VIF, variance inflation factor.

The a priori model, based on preselected variables, achieved an AUC of 0.80, and its discrimination and fit metrics lagged behind other methods. The partial correlation model explained the most variance (pseudo‐R ² = 0.28) but had a higher mean VIF, 3.28, indicating more multicollinearity. The PCA model reduced dimensionality to three principal components, but it explained less variance (pseudo‐R ² = 0.25) and exhibited poor fit as measured by the Hosmer–Lemeshow test (p < 0.001). Excluding the PCA model, variables that appeared in four to five of the models included median age, uninsured rate, fast‐food access, and elevation.

We compared logistic regression to a mixed‐effects model with Census Region as a random effect and an additional population‐weighted variant to account for clustering and regional differences. The comparison across the three approaches demonstrated consistent effects for the main model predictors, with no meaningful changes in the odds ratios (Table S2). In a second sensitivity analysis, we used a quantile regression model and the continuous prevalence of dementia diagnosis per 100,000. Using a backward selection process specific to quantile regression yielded nearly the same set of predictor variables as our main model. In addition, the pseudo‐R ² was the same, so we used the main model variables. We found no major shifts in predictor effects across the Q0.25 to Q0.75 range (Table S3), so we report the effects for a one standard deviation (SD) shift at Q0.50 in Figure S4. Because the sensitivity analysis findings were consistent with the main analysis, logistic regression was used for all subsequent analyses.

The main model incorporates household members, uninsured rate, diabetes prevalence, fast‐food access, smoking prevalence, white prevalence, and elevation as the predictors. We compared the main model to the components of the a priori model in Figure 3. When only including a priori demographic variables (age, education, black prevalence) the AUC was 0.76 (95% CI: 0.75 to 0.78), after adding socioeconomic variables (unemployment, uninsured, poverty) the AUC was 0.78 (95% CI: 0.76 to 0.80), and after adding medical comorbidities (diabetes, hypertension, stroke) the AUC was 0.80 (95% CI: 0.78 to 0.82). However, the main model's AUC is 0.84 (95% CI: 0.82 to 0.85), which is significantly better than any permutation of the a priori model (p < 0.001 for all comparisons). If you remove any variable from the main model, its discrimination and accuracy drop to the level of the a priori model or below.

FIGURE 3.

Receiver operating characteristic curves for four models illustrating improved prediction when using a stepwise approach to model building.

The main model results show varying associations of predictor variables on dementia diagnosis (Supplementary Table 2). Variables negatively associated with dementia diagnosis include number of household members, with an odds ratio of 0.50 (95% CI: 0.43 to 0.58); smoking prevalence, with an odds ratio of 0.46 (95% CI: 0.39 to 0.54); and average elevation, with an odds ratio of 0.55 (95% CI: 0.47 to 0.63), all indicating lower odds of dementia diagnosis for a positive SD shift. In contrast, positively associated variables include uninsured rate, with an odds ratio of 1.82 (95% CI: 1.57 to 2.12); diabetes prevalence, with an odds ratio of 5.34 (95% CI: 4.41 to 6.46); fast‐food access, with an odds ratio of 1.67 (95% CI: 1.50 to 1.86); and white prevalence, with an odds ratio of 3.84 (95% CI: 3.14 to 4.70), all suggesting higher odds of dementia diagnosis for a positive SD shift. Nonetheless, given the high degree of multicollinearity, the directionality of these effects may be confounded.

When we used the main model within the strata of the Census Regions, the AUC varied from 0.82 in the Midwest and South to 0.84 in the Northeast and 0.87 in the West (Table S4). If we used backward selection within each Census Region, it produced different sets of predictor variables as seen in Table S5. The result was an increase in AUC to 0.84 in the Midwest, 0.83 in the South, 0.86 in the Northeast, and 0.89 in the West. Variables in these regionally derived models that are not in the main model include Black prevalence, social vulnerability index, social deprivation index, average temperature, hospital admissions per 1000, food desert access, proximity to highway, stroke prevalence, poverty prevalence, fitness facilities, public assistance, physical inactivity, hearing loss, Medicare reimbursement, education, and crowded housing.

The regionally derived models demonstrated improved correct classification rates compared to the main model across all Census Regions. Overall, the main model correctly predicted the top tertile of dementia diagnosis in 77.5% of counties (Figure 4). In the Midwest, when using the regionally derived model, accuracy increased from 77.8% with the main model to 79.6%, an improvement of 1.8% (Table 4). The Northeast saw a more substantial gain, with correct classifications rising from 78.3% to 83.9%, an improvement of 5.6%. In the South, accuracy improved from 73.6% to 75.9%, a 2.3% gain, while the West experienced the smallest gain, rising from 89.2% to 90.9%, an improvement of 1.7% (Table 4). These improvements highlight the effectiveness of incorporating region‐specific predictors in enhancing classification accuracy, particularly in regions with initially lower performance.

FIGURE 4.

Prediction accuracy using main model, displayed for US counties on map of lower 48 states. Counties are categorized as Correct Prediction (blue) and Incorrect Prediction (orange) based on model performance.

TABLE 4.

Accurate versus inaccurate prediction of primary outcome shown by Census Region for main model (top) and region‐specific model (bottom).

Main model
Region	Total count	Correct	Incorrect
Midwest	1034	77.8%	22.2%
Northeast	217	78.3%	21.7%
South	1409	73.6%	26.4%
West	417	89.2%	10.8%

Region‐specific model
Region	Total Count	Correct	Incorrect
Midwest	1034	79.6%	20.4%
Northeast	217	83.9%	16.1%
South	1409	75.9%	24.1%
West	417	90.9%	9.1%

For the main model, the prediction characteristics included 1860 true negative, 468 false negative, 583 true positive, and 243 false positive classifications (see Figure S5 for a map of the prediction characteristics in the lower 48 United States). The mean values of the variables in the main model across these classifications is seen in Table S6.

The adjacency analysis included the 991 counties in the top tertile of dementia diagnosis prevalence compared to 1350 unique adjacent counties with lower prevalence. The counties in the top tertile of dementia diagnosis prevalence had higher diabetes prevalence (mean difference 0.53 SD, p < 0.001), greater fast‐food access (mean difference 0.18 SD, p < 0.001), reduced household members (mean difference −0.37 SD, p < 0.001), and higher smoking prevalence (mean difference 0.24 SD, p < 0.001) compared to their immediate neighbors. Counties with higher dementia diagnosis prevalence also had significantly lower elevation (mean difference −0.10 SD, p < 0.001). Differences in uninsured rate and White race prevalence were not statistically significant.

4. DISCUSSION

This study demonstrates that county‐level variation in demographic, medical, socioeconomic, and environmental factors can predict dementia diagnosis prevalence across the United States with high accuracy. Using a backward selection model, we identified seven predictors – higher diabetes prevalence, uninsured rate, fast‐food access, and White race prevalence and lower household size, smoking rate, and elevation – that achieved a high level of predictive accuracy (AUC = 0.84). Other models using different variable combinations performed similarly well, indicating that while dementia diagnosis prevalence can be predicted reliably, its drivers are multifaceted, interdependent, and difficult to isolate.

Interestingly, some predictors in the main model diverge from known biological mechanisms underlying dementia risk. Elevation, for example, showed a consistent negative association with dementia diagnosis prevalence in our analysis. This could suggest that higher elevation is linked to lower dementia risk, though equally likely is that it represents unmeasured confounding factors associated with higher‐elevation areas, such as healthier lifestyles, greater access to outdoor activities, or regional healthcare practices. Similarly, smoking prevalence was negatively associated with dementia, which contrasts with its well‐established detrimental effects on vascular health and cognitive outcomes. These paradoxical associations likely reflect contextual influences such as socioeconomic conditions or regional differences in healthcare access, which may confound the true effect of smoking. ⁴² , ⁴³ In this way, predictors in our models may function as “place‐based biomarkers,” serving as proxies for broader unmeasured factors that influence dementia diagnosis prevalence. Recognizing these nuances is essential for interpreting model results and applying them in public health contexts. While these predictors may not directly inform neurologic interventions, they can help identify high‐burden areas and guide tailored public health strategies, even when the underlying mechanisms are not fully understood.

We found the highest rates of dementia diagnosis prevalence in the Southern Census Region, which is consistent with prior research. ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ , ⁵² While early‐life environmental factors, such as being born in the South, are associated with higher dementia risk, particularly among Black and less‐educated individuals, more recent research highlights the critical influence of later‐life environments. ⁵³ , ⁵⁴ , ⁵⁵ Factors like state‐specific healthcare policies, socioeconomic conditions, and regional health behaviors appear to play a significant role in dementia diagnosis prevalence. ⁵⁶ , ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰

Our univariate findings confirmed expected associations with dementia diagnosis prevalence across key variables such as median age, household income, prevalence of disability, hearing loss, inactivity, poverty, hypertension, CHD, stroke, and diabetes. ⁹ , ⁵⁶ , ⁵⁹ , ⁶⁰ , ⁶¹ However, the multivariable models did not incorporate all of these variables simultaneously due to their overlapping contributions to the variance in dementia diagnosis. For instance, hypertension, diabetes, and stroke are closely interconnected through shared vascular mechanisms, while income and poverty represent similar dimensions of socioeconomic disadvantage. This process underscores the complementary roles of univariate and multivariable analyses: The former identifies broad patterns and associations, while the latter refines these insights to prioritize the most robust and interpretable predictors.

The findings of the adjacency analysis demonstrate that the risk factors identified in our main model explain dementia prevalence differences not only regionally but even between immediately neighboring counties, supporting the robustness of these associations at the county‐level geographic scale. However, this adjacency comparison included 2341 of the 3076 counties in our analysis (76%), indicating that the extensive geographic connectivity between counties means this spatial restriction provides limited additional discrimination beyond our primary model.

We used the backward selection set of variables for our main model because they achieved the best balance between parsimony, goodness of fit, and predictive accuracy. Median age was excluded from the final model, as its influence was already accounted for by the included variables and its addition did not improve model performance. This finding further demonstrates that dementia burden can be effectively predicted using a wide range of variable combinations. However, capturing additional variance may require the integration of granular, localized data. We did find that region‐specific models improved the predictive accuracy, highlighting the utility of localized predictors, such as fast‐food access, as targets for interventions in high‐burden regions to reduce dementia risk.

Our study is novel in its application of nationally representative datasets, which include over 61 million Medicare beneficiaries from diverse geographic and demographic backgrounds. The dataset integrates information on both FFS and Medicare Advantage enrollees, providing a uniquely comprehensive picture of dementia diagnosis prevalence across the United States. By aggregating county‐level data and incorporating diverse predictors from demographic, socioeconomic, healthcare, behavioral, and environmental domains, this analysis acknowledges the multidimensional nature of dementia risk. The hierarchical classification of the dementia diagnosis outcome, validated through prescription data and a 3‐year lookback period, adds further robustness to the study.

Our findings also suggest that predictive models must incorporate geographic variability and local data to improve accuracy and practical utility. While our main model performs well using a parsimonious set of predictors, we showed that its applicability was enhanced by incorporating region‐specific adjustments. For example, geographic variation in healthcare access, local socioeconomic conditions, and culturally specific health behaviors can significantly influence dementia diagnosis prevalence but may not be fully captured by national‐level predictors.

Supplementing national models with local measurements of dementia burden, such as those derived from studies like the Health and Retirement Study (HRS), ⁶² , ⁶³ and weighting subsequent predictions based on these data could further enhance the precision of estimates and improve the relevance of the models for regional interventions. This approach could help identify high‐priority areas for targeted dementia prevention efforts and potentially measure their effect over time.

Advances in diagnostic technology may substantially alter dementia diagnosis prevalence in the coming years. The anticipated US Food and Drug Administration approval of low‐cost plasma biomarkers for Alzheimer's disease, such as tau phosphorylated at threonine 217/amyloid beta 1‐42 ratios, could broaden access to accurate diagnosis beyond specialized centers. By reducing reliance on costly amyloid positron emission tomography imaging or cerebrospinal fluid testing, these biomarkers have the potential to democratize diagnostic capacity across counties and mitigate some disparities linked to specialist availability. Future geographic modeling should reassess predictors after widespread biomarker adoption to determine whether these tools reduce existing diagnostic disparities.

4.1. Limitations

Despite the strengths of this study, several limitations warrant consideration. First, the reliance on Medicare claims data means we are reporting the dementia diagnosis prevalence, not the prevalence of dementia. Although the two are highly related, the DDH may underestimate the true prevalence of dementia by excluding undiagnosed cases or individuals not enrolled in Medicare. Furthermore, while the DDH hierarchical classification algorithm is validated and robust, it may not capture subtle regional differences in diagnostic practices or criteria. Counties with more ready access to academic hospitals also likely have the means of obtaining appropriate cognitive screening and follow‐up neurocognitive assessment, including seeing the corresponding memory providers and undergoing appropriate clinical work up such as neuroimaging. For example, as of 2024, there are 504 United Council for Neurologic Subspecialties‐trained behavioral neurologists/neuropsychiatrists in the United States. ⁶⁴ The majority of practicing behavioral neurologists/neuropsychiatrists practice in academic settings, which can certainly affect a timely diagnosis of dementia for counties with less specialized healthcare access (Figure 1). Although the percentage of the population who are active neurologists was included as part of available county‐level variables, the number of memory diagnostic providers, which can include behavioral neurologists, neuropsychiatrists, and geriatric psychiatrists, would represent a separate pool of physicians, not specifically provided by the DDH. The regionally specific predictions regarding dementia diagnosis, with the Northeastern and West having higher accurate prediction compared with the South and Midwest counterparts (Figure 3 and Table 4), might also highlight that having closer and timelier access to the right clinical diagnostic pathway might play a sizeable role in dementia diagnosis.

Second, our analysis was limited by the availability, reliability, and granularity of the exposure data. Although the dataset is extensive, important predictors such as early‐life exposures or unmeasured social determinants of health may remain unaccounted for. In addition, many of the medical variables are self‐reported (e.g., hypertension, diabetes), which introduces recall bias.

Third, while our model highlights the potential of region‐specific adjustments, the integration of localized data into predictive frameworks remains a challenge. Obtaining reliable local data, validating it, and effectively weighting it within national models would require substantial additional resources and coordination. We also do not know how well the models would perform out of sample, for example, in different years of data.

Finally, using the top tertile of dementia diagnosis prevalence as the primary outcome, while clinically meaningful, may oversimplify the continuum of dementia diagnosis risk and reduce sensitivity to subtle variations in prevalence. Future research should incorporate additional years of data, explore alternative modeling approaches, and consider integrating survey and community‐based longitudinal exposure data to cross‐validate these associations.

5. CONCLUSION

The burden of dementia in the United States is shaped by a complex interplay of demographic, socioeconomic, and biological factors. The extensive collinearity of exposure variables, regional specificity of predictive exposures, and unexpected associations observed in our analysis emphasize the need for further investigation. Future research should integrate diverse data sources, including localized and individual‐level measures, to refine predictive models and enhance their applicability for targeted interventions. Ultimately, a multidimensional approach that accounts for both structural and individual‐level determinants will be essential in mitigating the burden of dementia and advancing efforts toward prevention and early intervention.

CONFLICT OF INTEREST STATEMENT

A. de Havenon has received research funding from the AAN, consulting fees from Novo Nordisk, and royalty fees from UpToDate and has equity in TitinKM and Certus. L. Littig reports no disclosures. G. Falcone reports no disclosures. R. Sharma reports no disclosures. A. Fesharaki reports no disclosures. S. Yaghi reports no disclosures. E. Calvario reports no disclosures. J. Rosand reports sponsored research support from the American Heart Association, National Institutes of Health, and consulting for the National Football League and Eli Lilly. K.N. Sheth reports compensation from Sense and Zoll for data and safety monitoring services, Cerevasc for consulting services, Rhaeos for consulting services, Certus for consultant services, and a patent pending for Stroke wearables licensed to Alva Health. C.D. Anderson reports sponsored research support from the American Heart Association, Bayer AG, and Massachusetts General Hospital and consulting for ApoPharma.

Consent Statement

All data were acquired from de‐identified existing datasets, so no consent statements were required.

Supporting information

Supporting information

Supporting information

de Havenon A, Littig L, Falcone GJ, et al. Unraveling disparities in county‐level dementia diagnosis prevalence across the United States. Alzheimer's Dement. 2025;21:e70954. 10.1002/alz.70954