A dynamic microsimulation method for estimating dementia costs in the United States

Abstract

INTRODUCTION

Addressing dementia's substantial economic burden on individuals, families, health systems, governments, and payers, requires integrating national data and advanced analytic methods.

METHODS

We described data sources, measures, and simulation methods of a dynamic microsimulation model developed for quantifying dementia costs in the United States and for assessing impact of innovation and change over time. Model internal validation was assessed.

RESULTS

Using nationally representative Health and Retirement Study data on adults over age 50, Medicare and other national data, we produced estimates of individual‐level health and economic outcomes over time. Estimates and simulation enabled measurement of population‐level dementia prevalence, medical and long‐term care costs, unpaid caregiving valuation, quality‐of‐life, and earnings for persons living with dementia and care partners. Counterfactual scenarios quantified dementia‐related quality‐of‐life and earnings losses. Simulated outcomes aligned with observed data.

CONCLUSIONS

Transparent methods support replication, extension, and informed use of dementia cost estimates in research and policy.

Highlights

• Developed a dynamic microsimulation model to estimate United Statessocietal dementia costs

• Integrated national data on health care, caregiving, quality of life, and earnings

• Modeled individual life‐course trajectories to produce population cost estimates

• Used counterfactuals to quantify unobserved costs of dementia

• Provided methodological details for transparency and assessed model validity

1. BACKGROUND

The economic impact of dementia in the United States is increasing rapidly as the population ages. This burden extends well beyond direct medical expenditures to include long‐term services and supports, substantial unpaid caregiving by family members and friends, losses in earnings and productivity, and significant declines in quality of life for people living with dementia and their care partners. ¹ , ² , ³ Together, these effects impose a growing societal burden that affects households, health systems, and public and private payers.

Understanding the magnitude and distribution of these costs—who bears them, in what forms, and over what time horizons—is essential for informed decision making about resource allocation and policy priorities. As prevention strategies, treatments, care, and health policies evolve, both the level and composition of dementia‐related costs will change, as will the balance of costs borne by individuals, families, and society. Approaches that focus narrowly on costs attributable to dementia often fail to capture these broader impacts, particularly those arising from changes in long‐term care use, family caregiving, and quality of life.

To address these limitations, the United States Cost of Dementia Project team developed the United States Cost of Dementia Model (USCDM), a dynamic microsimulation model designed to estimate the societal costs associated with dementia in the United States. The USCDM quantifies the real‐world resource burden on families, health systems, and society by capturing how dementia shapes medical care, long‐term care, and family caregiving in ways that cannot be easily disentangled from co‐occurring conditions. This comprehensive approach enables meaningful comparisons of disease burden across conditions and supports better‐informed targeting of resources. The USCDM was predated by the Future Elderly Model, which was utilized to study many research questions around the aging population. It was extensively validated for outcomes related to quantity and quality of life, including dementia. ⁴ , ⁵

The model uses nationally representative data from sources such as the Health and Retirement Study (HRS) and the Centers for Medicare & Medicaid Services to generate estimates for the U.S. population. The dynamic microsimulation model provides the methodological tools for assessing how dementia prevalence and costs today and in future years will change with, for example, changes in disease risk, policies to support caregivers in the workforce, treatment innovation, and improved dementia care. These may be actual changes or hypothetical. Additionally, the model is used retrospectively to quantify what health and economic outcomes could be in the absence of dementia, enabling measurement of quality‐of‐life and earnings losses, thereby capturing the long‐run, cumulative consequences of dementia on health and economic outcomes.

The goal of this paper is to document the USCDM as used to estimate the societal cost of dementia in 2026. While a companion paper presents the estimates of the annual cost of dementia in the United States for 2026, this methods paper provides a detailed description of the model structure, data sources, assumptions, valuation approaches underlying those estimates and results of validation analyses. By making the methods explicit, this paper aims to enhance transparency, facilitate replication and extension, and support appropriate interpretation and use of dementia cost estimates in research and policy analysis.

RESEARCH IN CONTEXT

1. Systematic review: Reviewed studies of dementia's economic impact. Prior work narrowly defined costs, varied in methods, and included limited detail to support replication and extension of dementia cost estimates.

2. Interpretation: This study described a dynamic microsimulation model that quantified comprehensive societal costs of dementia by capturing the complex, interconnected ways dementia costs evolve over time. The model simulated what outcomes could be in the absence of dementia enabling measurement of quality‐of‐life and earnings losses along with medical and long‐term care costs and valuation of caregiving. Detailed and transparent methods description support replication, extension, and informed use of estimates in research and policy.

3. Future directions: The flexible model can incorporate new data as it becomes available and be used to evaluate how dementia costs may change under alternative prevention strategies, treatment and care innovations, and policy interventions, supporting informed decision making and future research on dementia's societal burden.

2. METHODS

2.1. Model overview

The USCDM is a dynamic microsimulation model. The model uses nationally representative data from the HRS complemented with other national data sources to generate dementia cost estimates for the U.S. population aged 51 and older. The HRS interviewed both respondents and their spouses, and they report on their families providing rich information on family financial and care‐related resources. The USCDM estimates dynamic models of health and economic outcomes using longitudinal data from the 1998 to 2020 waves of HRS. The modeling of individuals who are followed longitudinally allows for substantial heterogeneity in behavior. The microsimulation begins with HRS respondents aged 51 years or older in year 2010. The USCDM has a 2‐year cycle that mimics the frequency of the underlying survey data. At each two‐year timestep, individual characteristics and outcomes are updated based on predictions from the models. Individual‐level estimates are aggregated to the U.S. population using survey weights. USCDM produces annual estimates of dementia prevalence, medical and long‐term care costs, valuation of hours of care received family and others, loss of quality of life and earnings for persons living with dementia and separately for their care partners. USCDM is a flexible and policy‐relevant tool for evaluating how prevention strategies, treatment and care delivery innovation, and policy reform may influence future dementia‐related costs. The USCDM study was reviewed and approved by the University of Southern California Institutional Review Board under expedited review.

2.2. Model architecture and simulation framework

2.2.1. Overview of core components of USCDM

The USCDM has three core components: the Transition Module is the core of USCDM; the Initial Conditions Module replenishes the simulation population; the Policy Outcomes Module aggregates population health care spending. The Transition Module calculates the probabilities of various health states (e.g., cognitive impairment, incidence of diabetes) and financial outcomes (e.g., work, earnings) in a Markov framework. The module takes as inputs health behaviors (e.g. smoking), demographic (e.g., age) and socioeconomic risk factors (e.g., education), along with lagged health and financial states. This allows for heterogeneity across individuals. The transition probabilities are estimated from the longitudinal data in the HRS. The Initial Conditions Module jointly predicts the economic and health outcomes of replenishing cohorts of 51‐ and 52‐year‐olds. This module takes in data from the HRS and trends calculated from the National Health Interview Survey (NHIS) and the National Health and Nutrition Examination Study (NHANES) to “generate” cohorts as the simulation proceeds. Additionally, the sample weights of these replenishing individuals are adjusted by sex and race and ethnicity to ensure that the simulation remains nationally representative of the 51 and older population in the United States. The Policy Outcomes Module aggregates projections of individual‐level outcomes into policy outcomes such as total medical care costs and by payer. This component takes account of public and private program rules (e.g., eligibility).

2.2.2. Transition module

We modeled transitions for a broad set of outcomes observed over mid‐life and later life (Table 1). Because the HRS sampled individuals aged 51 and older, respondents entered the data at different ages and were followed for different lengths of time. As a result, each individual contributed observations over a personal sequence of ages, beginning at the first age at which they were observed and ending at their last observed age. This created an unbalanced panel covering ages 51 and above, with outcomes observed repeatedly for each respondent. Different birth cohorts were represented in the HRS and the study was refreshed with new cohorts of 51‐ to 56‐year‐olds periodically over time, thus transition model estimates do not rely on a single age cohort. The richness of the data on each individual accounted for cohort differences, and survey weights were adjusted to represent the U.S. population. We assumed that the relationships measured in the transitions for predicting outcomes were consistent across cohorts over time conditional on a rich set of covariates.

TABLE 1.

USCDM key health and economic variables by module.

Economic	Health
Initial Conditions Module
Employment	Cognitive status
Wealth	Activities of daily living
Earnings	Instrumental activities of daily living
	Diabetes
	Hypertension
	Heart disease
	Self‐reported health
	BMI
	Smoking status
	Functional status
Transition Module
Employment (full‐time, part‐time, not working)	Cognitive status
Earnings	Activities of daily living
Household wealth	Instrumental activities of daily living
Health insurance	Diabetes
Medicare Part B enrollment	Hypertension
Medicare Part D enrollment	Heart disease
Help hours received from family and friends	Stroke
Help hours provided to parents	Cancer
	Lung disease
	Congestive heart failure
	BMI
	Smoking status
	Nursing home
	Mortality
	Depressive symptoms (CES‐D)
	Parent of respondent mortality
	Parent of respondent has dementia
	Parent of respondent co‐residency
Policy Module
Out‐of‐pocket medical and LTC expenditures	Quality‐adjusted life‐years
Private Insurer medical expenditures
Medicare Part A expenditures
Medicare Part B expenditures
Medicare Part D expenditures
Medicaid expenditures

Note: Other economic outcomes are available for use with USCDM including those related to pension coverage and wealth, social security, disability, SSI receipt, taxes, other government transfers. Details on data and methods relevant to government revenues and expenditures can be accessed at https://schaeffer.usc.edu/wp‐content/uploads/2025/02/FEM‐Technical‐Document.pdf.

Abbreviations: BMI, body mass index; CES‐D, Center for Epidemiological Studies‐Depression; LTC, long‐term care; SSI, Supplemental Security Income; USCDM, United States Cost of Dementia Model.

We modeled absorbing and non‐absorbing types of outcomes. Absorbing outcomes were those assumed to persist permanently once they occur, for example, heart disease. The risk of experiencing a new absorbing outcome depended on both time‐invariant and ‐variant factors. Time‐invariant factors were fixed individual characteristics (such as sex or education) and health conditions that were already present before the respondent first enters the study, such as an indicator of ever having a diabetes diagnosis before age 51. Time‐varying factors included other outcomes that were reported in earlier periods, such as prior diagnoses of different conditions. This structure allowed earlier health events to affect the risk of developing additional conditions later in life. Unobserved time‐varying shocks were assumed to be normally distributed and independent across diseases.

We modeled three types of non‐absorbing outcomes: binary, ordered, and censored. Binary, non‐absorbing outcomes could change over time, and current status may depend on past status capturing state dependence. Whether an individual lived in a nursing home is an example. Ordered outcomes were modeled as arising from an underlying continuous measure, with thresholds determining the observed categories. As with binary outcomes, we allowed past outcomes to influence current values. Cognitive status representing no cognitive impairment, mild cognitive impairment, and dementia was modeled as an ordered non‐absorbing state. Censored outcomes included earnings and financial wealth. Earnings were only observed when individuals were working, and wealth included a substantial number of zero and negative values. Negative wealth values indicate a household with debts exceeding assets. To address this, we used two‐part models: first modeling whether the outcome was observed and then modeling its level conditional on being observed. To keep the model realistic and tractable, we restricted how outcomes can feed back into one another over time. These pathways were developed in consultation with clinicians and scientists, beginning with prior model versions. Table 2 displays which past outcomes are allowed to influence future transitions.

TABLE 2.

Health pathways.

At time t+1→	BMI	Smoking	Cancer	Diabetes	Heart disease	Hypertension	Lung disease	Stroke	CHF	Heart attack	Dementia	ADLs	IADLs	CES‐D	Mortality
BMI	✔	✔	✔	✔	✔	✔	✔	✔		✔		✔	✔	✔
Smoking	✔	✔	✔	✔	✔	✔	✔	✔	✔	✔	✔	✔	✔	✔	✔
Cancer	✔	✔						✔			✔	✔	✔		✔
Diabetes	✔	✔			✔	✔		✔	✔	✔	✔	✔	✔		✔
Heart disease	✔	✔		✔		✔		✔	✔		✔	✔	✔		✔
Hypertension	✔	✔		✔	✔			✔	✔	✔	✔	✔	✔		✔
Lung disease	✔	✔									✔	✔	✔		✔
Stroke	✔	✔		✔	✔	✔				✔	✔	✔	✔		✔
CHF															✔
Heart attack	✔	✔		✔		✔		✔		✔	✔	✔	✔	✔	✔
Dementia (cognitive state)											✔		✔		✔
ADLs	✔	✔										✔	✔	✔	✔
IADLs	✔	✔										✔	✔	✔	✔

Notes: ✔ indicates that a direct lagged effect of the condition in Column 1 is included in the corresponding transition equation of the health and functional outcomes in top row. All transition models additionally adjust for a common set of demographic and socioeconomic covariates, including age, sex, race/ethnicity, educational attainment, marital status, labor‐force status, and baseline health‐state indicators. The parameterization of covariate effects varies across models.

Abbreviations: ADL, activities of daily living; BMI, body mass index; CESD, Center for Epidemiological Studies‐Depression; CHF, congestive heart failure; IADL, instrumental activities of daily living;

Model parameters were estimated using ordinary least squares, probit, and ordered probit models, with some exceptions for earnings and wealth. Given our assumptions that unobserved shocks were normally distributed and independent, the probability of each individual's observed sequence of outcomes over time can be written as a product of simple probabilities. Because outcome processes were assumed to be independent once we conditioned on initial characteristics, the overall likelihood was constructed by multiplying probabilities across outcomes and time periods. To address the skewed distributions of earnings and wealth (and negative wealth values), we applied a flexible transformation—a generalized inverse hyperbolic sine transformation—which allowed us to handle skewness while retaining observations with zero or negative values.

2.2.3. Initial conditions module

The Initial Conditions Module jointly predicted economic and health outcomes for new cohorts of 51‐ and 52‐year‐olds. Projections combined two sources of information: (1) observed time trends for individual outcomes among younger cohorts and (2) the cross‐sectional relationships among outcomes observed at age 51 in the HRS. At any calendar time, individuals were characterized by a vector of outcomes (e.g. health conditions), which together defined a joint population distribution. This joint distribution captured both the prevalence of each outcome and how outcomes co‐occur. In principle, one would observe how this full distribution evolves over time, but available evidence typically consisted of trends for single outcomes considered separately. For each outcome, we observed marginal distributions over time, for example, repeated cross‐sections from the NHIS show rising diabetes prevalence among 51‐year‐olds. These data described changes in average levels but provided no information on how outcomes were jointly related as those trends unfolded. By contrast, the HRS provided a snapshot of the joint distribution of outcomes at a given point in time (e.g., age‐51 respondents in 1992), capturing correlations among outcomes. Our approach combined these sources by allowing marginal distributions of individual outcomes to evolve according to observed trends, while assuming that the relationships among outcomes remained stable over time. Thus, average levels may change, but patterns of co‐occurrence were preserved. This strategy allowed us to project realistic initial conditions for new cohorts that reflected known trends while maintaining empirically grounded associations among outcomes.

2.2.4. Implementation

Figure 1 illustrates the simulation framework. To align with the biennial structure of the HRS data used to estimate transition models, the USCDM advanced in 2‐year increments, with each step ending on July 1 to facilitate integration with Social Security Administration population demographic projections. The simulation was initialized in 2010 using a population representative of U.S. adults aged 51 years and older, constructed from the HRS.

FIGURE 1.

Temporal structure of the U.S. cost of dementia model.

At each 2‐year step, the USCDM applied transition models for mortality, cognitive status, other health conditions, employment, earnings, and nursing home residence. Individual outcomes were simulated using Monte Carlo draws based on predicted transition probabilities. For each outcome, the model compared an individual's predicted probability of experiencing a transition over the next 2‐year period to a random draw to determine whether the transition occurs. For example, an individual with a 5% predicted probability of developing diabetes during a given time step will be assigned diabetes in approximately 5% of simulated realizations. The population was then updated to reflect net immigration using U.S. Census Bureau projections stratified by age, sex, and race, and new cohorts of 51‐ and 52‐year‐olds were added in numbers and compositions consistent with Census estimates.

After transitions and cohort replenishment, cross‐sectional models were used to estimate medical costs, and summary outcomes were computed. Medical cost calculations included individuals who died during the interval to capture end‐of‐life expenditures. This sequence—from population initialization through transitions, cost estimation, and cohort replenishment—was repeated for each 2‐year cycle, as illustrated in Figure 2. To reduce stochastic variability, the simulation was replicated 50 times, and results were averaged across runs. Uncertainty in summary outcomes was quantified using nonparametric bootstrapping with 6,000 Monte Carlo simulations.

FIGURE 2.

U.S. cost of dementia model simulation flow.

The USCDM required assumptions regarding growth in wages, retirement age, real medical cost growth (defined as growth in medical spending above general inflation), interest rates, cost‐of‐living adjustments, the consumer price index, and the Social Security Administration's substantial gainful activity threshold. Default assumptions were drawn from the 2010 Social Security Trustees ⁶ intermediate‐cost scenario and were specified in real 2010 dollars; results were converted to nominal dollars as needed post‐simulation. Transition and cross‐sectional models were estimated in Stata, including the incoming cohort model using the CMP package, ⁷ and the simulation engine was implemented in C++.

2.3. Data sources

The USCDM combined data from several large, nationally representative surveys. The primary data source was the HRS, which was used to define the baseline simulation population, generate replenishing cohorts, and estimate transition models. We supplemented the HRS with annual data on health care use and spending from the Medical Expenditure Panel Survey (MEPS) for the non‐Medicare population and the Medicare Current Beneficiary Survey (MCBS) for the Medicare population. The combination of datasets provided more comprehensive information than any single national survey.

The NHIS was used to project the prevalence of chronic conditions in replenishing cohorts, and the NHANES was used to project body mass index (BMI) in future cohorts of 51‐ and 52‐year‐olds. In addition, population counts and projections were obtained from the American Community Survey (ACS) and the U.S. Census Bureau. We describe these sources and data used for projections in more detail below.

2.3.1. HRS

The HRS, initiated in 1992, is a biennial, nationally representative longitudinal survey of approximately 40,000 U.S. adults aged 51 years and older. It was the primary longitudinal data source for modeling cognition and dementia, health conditions, functional limitations, health care and long‐term care use, caregiving, health‐related quality of life, and socioeconomic outcomes, including employment, earnings, wealth, and insurance eligibility. Biennial data from 1998 to 2020 were used to estimate transition models. The HRS replenishes the sample every 6 years by adding new cohorts aged 51–56, and when weighted, survey waves from 1998, 2004, 2010, 2016, and 2022 are representative of U.S. households with at least one member aged 51 or older. The HRS served as the host population for the simulations of individuals aged 51 and older. The study is sponsored by the National Institute on Aging (NIA; U01AG009740) and conducted by the University of Michigan. Analyses use harmonized measures from the RAND HRS Longitudinal File when available. ⁸

2.3.2. MEPS

The MEPS, initiated in 1996, is a nationally representative survey of the U.S. non‐institutionalized population, their medical providers, and employers. MEPS employs an overlapping 2‐year rotating panel design, with each panel interviewed five times over approximately 2.5 years. Each annual sample includes about 12,000 households (34,000 individuals), including roughly 4,500 adults aged 51–64. The MEPS Household Component collects detailed information on demographic characteristics, health conditions, medical service use, and sources of payment, supplemented with data from medical providers. Households are drawn from the prior year's NHIS, which excludes institutionalized individuals; accordingly, MEPS was used to estimate medical costs for the non‐institutionalized population. MEPS data from 2007–2010 were used to estimate health care utilization and spending models for individuals with dementia under age 65 who are not enrolled in Medicare through disability entitlement. Estimated relationships were applied to comparable individuals with dementia in the HRS using aligned socioeconomic measures. Year‐over‐year excess real growth in medical costs was applied in simulations beginning in 2010.

2.3.3. MCBS

The MCBS, initiated in 1991, is a nationally representative survey of Medicare beneficiaries, including older adults, individuals with disabilities, and those residing in institutional settings. Since 1994, the MCBS has used a rotating panel design in which beneficiaries are followed for up to 3 years and interviewed up to 12 times, regardless of residence or transitions between community and institutional care. Each year, a new panel of approximately 12,000 beneficiaries is introduced and an older panel is retired. Individuals under age 65 with disabilities and adults aged 85 and older are oversampled, and institutionalized respondents are interviewed by proxy. The MCBS collects detailed information on health conditions, health care use and expenditures, insurance coverage, and socioeconomic and demographic characteristics. Dementia is identified using self‐ or proxy‐reported physician diagnoses, supplemented with linked Medicare fee‐for‐service claims. MCBS data from 2007 to 2012 were used to estimate medical utilization, cost, and enrollment models for Medicare beneficiaries, which were applied to individuals with dementia in the HRS. As with MEPS‐based projections, year‐over‐year excess real growth in medical costs was applied in simulations beginning in 2010.

2.3.4. NHIS

The NHIS is a nationally representative, repeated cross‐sectional survey of the U.S. non‐institutionalized population conducted annually for several decades. NHIS collects individual‐level data on self‐reported memory problems, cognitive screening measures, chronic conditions, income, education, and demographic characteristics. Beginning in 1997, a subsample of adults was administered a comprehensive chronic conditions module. Survey content and question wording are closely aligned with the HRS. NHIS data from 1997 to 2010 were used to project the prevalence of chronic conditions among future cohorts entering the HRS at ages 51–52 by extrapolating observed trends from younger cohorts as they age into eligibility.

2.3.5. NHANES

The NHANES, conducted by the National Center for Health Statistics, is a nationally representative survey combining interviews, physical examinations, and laboratory assessments. NHANES provides objective health measures, including BMI, blood pressure, and biomarkers, alongside self‐reported health information. These data were used to estimate trends in BMI distributions, which inform the assignment of BMI for replenishing cohorts entering the simulation.

2.3.6. Supplemental data sources and projection assumptions

Additional data sources were used to project wages, medical costs, population demographics, migration, and mortality. Real wages were assumed to grow at 1.1% annually in the long run. Excess real growth in medical costs (growth above gross domestic product [GDP]) followed Centers for Medicare & Medicaid Services assumptions, declining linearly from 1.5% in 2004 to 1.0% in 2033; Affordable Care Act cost‐growth targets were incorporated as an optional cap. Near‐term GDP growth was based on Congressional Budget Office projections, with long‐run growth of 2.0% per year thereafter, consistent with Federal Old‐Age and Survivors Insurance and Federal Disability Insurance Trust Funds (OASDI) Trustees assumptions. Population counts and demographic trends were aligned using the American Community Survey, National Population Projections, and the Human Mortality Database. HRS sample weights were post‐stratified by five‐year age group, sex, and race to match 2010 American Community Survey population totals for both the host sample and replenishing cohorts. Weights for future cohorts were scaled using 2012 National Population Projections by sex and race. Mortality weights for decedents were similarly post‐stratified using the Human Mortality Database. Net migration and mortality trends were based on the Social Security Administration (SSA) Trustees intermediate‐cost scenario. All‐cause mortality adjustment factors were defined separately for individuals under and over age 65 using 2013 SSA Trustees projections, which were interpolated through 2090.

2.4. Outcomes

2.4.1. Dementia

At each simulation year, the USCDM quantified (1) the number of individuals living with all‐cause dementia and (2) the number of individuals with dementia residing in nursing homes or long‐term care facilities. Cognitive function of HRS respondents was assessed at each wave using an adapted version of the Telephone Interview for Cognitive Status (TICS). ⁹ Cognitive status was classified using the Langa–Weir algorithm, ¹⁰ which categorizes individuals as cognitively unimpaired, cognitively impaired without dementia, or cognitively impaired with dementia based on self‐ or proxy‐reported information. For self‐respondents, the 27‐point TICS score combined immediate and delayed word recall (0–20), serial sevens (0–5), and backward counting (0–2). Dementia was defined as scores of 0–6, cognitive impairment without dementia as 7–11, and normal cognition as 12–27. For proxy respondents, an 11‐point scale was constructed using instrumental activities of daily living (ADL) limitations, interviewer‐rated cognitive impairment, and proxy‐rated memory. Proxy scores of 6–11 indicated dementia, 3–5 cognitive impairment without dementia, and 0–2 normal cognition. Proxy and self‐respondent classifications were combined in the analytic sample.

Cognitive status transitions (no cognitive impairment, mild cognitive impairment, dementia) were estimated using ordered probit models adjusted for demographic characteristics, health behaviors, and health conditions. Covariates included age‐splines, sex, education, race/ethnicity, widowhood, and one‐wave lags of cognitive status, chronic conditions associated with dementia (diabetes, heart disease, hypertension, stroke), as well as other health conditions (lung disease, cancer), BMI (lag and at age 50), and smoking status (lag and at age 50). Models also included interactions between sex and race/ethnicity and between education and age.

2.4.2. Nursing home status

Residence in a nursing home or long‐term care facility was determined using respondent or proxy reports of living arrangements, based on the RAND HRS data and variable nhmliv. Transitions into nursing home residence were modeled using probit regressions including lagged nursing home status, sociodemographic characteristics, lagged and age‐50 health conditions, lagged cognitive impairment and dementia, lagged functional limitations, widowhood, wealth, calendar‐year trends, and interactions between sex and race/ethnicity and between education and age. Table 3 provides a summary of the outcomes affected by dementia.

TABLE 3.

Outcomes directly affected by dementia.

Person with dementia is:	Domain	Direct effects
Individual	Cognitive status	Subsequent cognitive status
	Mortality	Death
	Function	Instrumental activities of daily living
	Living arrangement	Nursing home residence
	Care received	Any spousal help
		Spousal help is full time
		Annual spousal help hours
		Any non‐spousal help
		Non‐spousal help is full‐time
		Annual non‐spousal help hours
	Health‐related quality of life	Health utility (HUI3 score)
	Mental health spillovers	Spouse depressive symptoms (CES‐D score)
	Labor supply	Currently working for pay
		Full‐time or part‐time work
	Prescription drugs	Any prescription drug use
		Prescription drug expenditures
	Medical spending	Medicare Part A expenditures
		Medicare Part B expenditures
		Medicaid medical expenditures
		Out‐of‐pocket medical expenditures
		Total medical expenditures
Individual's parent	Living arrangement	Mother lives with respondent
		Father lives with respondent
	Parental mortality	Mother's death
		Father's death
	Care provided to parent	Caregiver to mother
		Caregiver to father
		Annual caregiving hours provided to parents
	Mental health spillover	Respondent depressive symptoms (CES‐D score)

Abbreviations: CES‐D, Center for Epidemiological Studies‐Depression; HUI3, Health Utility Index 3.

2.4.3. Hours of unpaid care received

At each simulation year, the USCDM quantified (1) the number of individuals with dementia receiving informal care from a spouse, (2) the number receiving care from non‐spousal caregivers (e.g., children, other relatives, or friends), and (3) the total hours of unpaid caregiving received from family and others.

Care received was measured using self‐ and proxy‐reports of hours and days of assistance received from family and friends for ADL and monthly hours of help with instrumental activities of daily living (IADL), excluding paid care. Reported care hours were converted to annual hours and estimated separately for spousal and non‐spousal caregivers among respondents classified as having dementia.

Care hours were modeled using a three‐stage approach. The first stage estimated the probability of receiving any unpaid care from family or others. Conditional on receiving care, the second stage estimated whether care received was full‐time (> 6,570 hours per year summed across hours received from all caregivers) or part‐time. For individuals receiving full‐time care, caregiving hours were assigned as the mean observed hours within spousal and non‐spousal full‐time caregiver groups. For individuals receiving part‐time care, the third stage estimated annual caregiving hours. Covariates in all stages included recipient sex, age, race/ethnicity, education, marital status, number of children, chronic conditions, functional limitations, cognitive status (mild cognitive impairment or dementia), and lagged caregiving hours. All covariates were interacted with cognitive status.

2.4.4. Cost components

At each simulation year, the USCDM quantified the dollar value of unpaid caregiving, total medical and long‐term care spending by payer (Medicare, Medicaid, other public or private sources, and out‐of‐pocket), quality‐of‐life losses for individuals living with dementia, quality‐of‐life losses for unpaid caregivers (spousal and non‐spousal), and earnings losses for both individuals with dementia and their family caregivers.

2.4.5. Dollar value of unpaid care received from family and others

The monetary value of unpaid caregiving was estimated using annual hours of care received by individuals with dementia and two valuation approaches. The replacement‐cost method valued caregiving hours at $34.92 per hour, reflecting the average cost of hiring a paid caregiver. ¹¹ The opportunity‐cost method valued caregiving time using caregivers’ potential wages. Hourly wage rates were derived from HRS‐reported wages at age 51 across all survey waves; when unavailable, wage rates were predicted using an estimated wage model. In the opportunity cost approach, caregiving hours were valued at $30.42 per hour (2025 dollars) for spousal caregivers and $34.62 per hour for non‐spousal caregivers, including adult children and other relatives.

2.4.6. Medical and nursing home spending

Total medical and long‐term care spending for individuals living with dementia was quantified overall and by payer, including Medicare (Parts A, B, and D), Medicaid, other payers, and out‐of‐pocket spending. Spending models were estimated using pooled data from the MCBS for Medicare‐eligible individuals and from the MEPS for those not eligible for Medicare. All spending was inflated to constant dollars using the medical component of the Consumer Price Index. Because MEPS and MCBS are known to understate medical expenditures, adjustment factors were applied so that predicted per‐capita spending matched aggregate totals from the National Health Expenditure Accounts (NHEA).

For the Medicare‐eligible population, annual medical spending was estimated using ordinary least squares models based on MCBS data. Models included demographic and socioeconomic characteristics of age, race and ethnicity, education, and marital status. Dementia and other chronic conditions (cancer, diabetes, hypertension, heart disease, lung disease, stroke, acute cardiac events) were included using indicators distinguishing the incident year of diagnosis from subsequent years. Additional controls included severe functional impairment (three or more ADLs), nursing home residence, Medicare eligibility via disability insurance, and death during the year, with interactions allowing end‐of‐life spending to vary by age, condition, nursing home residence, and disability status. Selected condition interactions were included to capture nonadditive effects of comorbidities. Medical costs for nursing home residents were aggregated to estimate total medical spending for individuals with dementia residing in nursing homes.

Once eligible for Medicare due to age or 2 years of Social Security Disability Insurance (SSDI) receipt, universal enrollment in Medicare Part A was assumed. Medicare Part B enrollment was estimated using weighted probit models and pooled MCBS data, with separate models for new enrollees and existing beneficiaries. To correct for overstatement of Part B enrollment in MCBS, a premium elasticity of −0.10 was applied and calibrated to observed enrollment trends. Part B premiums were calculated using prior‐period average costs and Affordable Care Act means‐testing thresholds and included in models of Part B enrollment. Medicare Part D enrollment and costs were modeled using MCBS data beginning in 2006. Enrollment probabilities were estimated using probit models, with time‐varying adjustments to account for underreporting and projected enrollment growth. Conditional on enrollment, Part D costs, using capitated payment amounts, were estimated using ordinary least squares models and demographic, health and function variables as reported above. Estimated Part D spending was combined with Parts A and B to calculate total Medicare expenditures, after which medical cost growth assumptions were applied.

The USCDM does not currently model enrollment in Medicare Part C. In the MCBS data, capitated payments are used to estimate total medical spending for individuals enrolled in Part C. Within the simulation, however, all Medicare enrollees are assigned costs as if they were enrolled in fee‐for‐service Medicare. Health care use of persons with dementia in Medicare Advantage (Part C) and traditional Medicare is similar with the exception of the months immediately around diagnosis. ¹² Alignment of MCBS costs to national NHEA estimates helps mitigate any bias in applying fee‐for‐service estimates to all.

For the non–Medicare‐eligible population, annual medical expenditures were estimated using ordinary least squares models based on MEPS data. Models included demographic characteristics, education, marital status, earnings, chronic conditions, and interactions allowing spending patterns to vary by sex across demographic groups. Costs are estimated for all respondents at each time step in the simulation and reported separately for HRS respondents living with dementia in nursing homes and persons with dementia not living in nursing homes.

2.4.7. Quality of life loss due to dementia for persons living with dementia

Quality‐adjusted life‐years (QALYs) are based on Health Utility Index 3 (HUI3) measure assessed for HRS respondents in 2000. The HUI3 measures eight dimensions of health including vision, hearing, speech, ambulation, dexterity, emotion, pain, and cognition. It was mapped to a single score ranging from −0.36 to 1 where 0 represents death, negative scores are worse than death, and 1 is perfect health. ¹³ We used these utility scores to compute QALYs. We estimated OLS model of HUI3 scores with the following covariates: marital status, ADL, IADL, cognitive status, health conditions, smoker status and depressive symptoms. To determine the loss in quality of life associated with living with dementia, we simulated a counterfactual scenario in which people with dementia in 2010 instead had mild‐cognitive impairment and, with the exception of death, remained with mild cognitive impairment through 2026. The difference in total number of QALYs in this counterfactual scenario with the status quo scenario in 2026 is multiplied by $150,000, the value for the loss of one QALY.

2.4.8. Quality of life loss for dementia caregivers

We estimated the effect of being a caregiver and hours of caregiving on depressive symptoms measured using the eight‐item Center for Epidemiological Studies Depression scale (CES‐D). We estimated the effect of depressive symptoms on health utility (HUI3) and used the utility scores to compute QALYs. We estimated QALYs for respondents who reported providing care for living parents or parents‐in‐law, or for spouses living with dementia. Details follow.

From the RAND HRS family data files, ⁸ we used information on the amount of care provided to a parent (100 or more hours in a year), as well as characteristics of the parent (e.g. age, education) and whether the parent is co‐residing with the respondent, owns a home, and has Alzheimer's disease or dementia. We included these characteristics of the parents in models predicting whether a respondent provided care for their parents and, subsequently the amount of care provided. The models also included the respondents’ age, race/ethnicity, education, widowhood, wealth, earnings, functional status, number of siblings, and chronic conditions, and were estimated separately for male and female caregivers.

We estimated CES‐D scores as an ordered probit. Key covariates were whether any parents had dementia, whether a spouse had dementia, whether the respondent was a caregiver to spouse or parents, and the number of hours per year spent providing care to parents or spouses. Other covariates included respondents’ age, race/ethnicity, sex, education, ADLs, IADLs, widowhood, health conditions at age 50 (e.g. diabetes, heart disease), health risk factors and behaviors (smoking, BMI).

In order to estimate the loss in QALYs due to caring for parents with dementia, we simulated a counterfactual scenario such that parents did not have and do not acquire dementia, and respondents do not provide dementia care to parents (though non‐dementia care is still provided). The QALY loss was the difference in total QALYs between this scenario and the status quo scenario where parents acquired dementia and may be cared for by their children. Similarly, in the counterfactual scenario for respondents who had a spouse with dementia, we simulated the impact on QALYs if a spouse did not have dementia. The estimated QALY losses are calculated as the difference in total QALYs between each counterfactual and the status quo. The total QALY loss associated with providing care to parents was divided by number of persons providing care to parents to quantify an average per‐capita QALY loss. Similarly, the total QALY loss associated with providing care to a spouse was divided by number providing care to a spouse for an average per‐capita QALY loss. The average per‐capita QALY loss associated with care to parents was multiplied by the total number of non‐spousal caregivers, and the average per‐capita QALY loss associated with care to a spouse was multiplied by the total number of caregivers to a spouse and finally summed for quality of life loss for dementia caregivers.

2.4.9. Earning loss for persons living with dementia

Work and full‐time status were modeled using probit regressions that included sex, race/ethnicity, education, age and retirement timing, mild cognitive impairment, dementia and other current and lagged health conditions, functional limitations, prior labor force participation, program and benefit claiming, and caregiving to parents with and without dementia, including 2‐year lags. All covariates were interacted with indicators for mild cognitive impairment and dementia.

Earnings were measured using RAND HRS data and included wages, bonuses, overtime pay, commissions, tips, earnings from second jobs, military reserve income, and income from professional practice or trade. Earnings were modeled using linear specifications with an inverse hyperbolic sine transformation for full‐time workers due to the long right tail and ordinary least squares models for part‐time workers. All earnings equations used a common set of covariates, including age, sex, race/ethnicity, education, health conditions and health behaviors at age 50, and 2‐year lags of health conditions, functional limitations, widowhood, earnings, and benefit claiming.

Earnings losses associated with dementia were estimated using a counterfactual simulation in which individuals with dementia in 2010 were assumed to have mild cognitive impairment and remain unchanged through 2026. This was chosen as a more conservative approach compared to assuming no cognitive impairment. Counterfactual earnings were generated by allowing cognitive status to affect employment and full‐time status, which in turn determined the applicable earnings equation. Earnings losses were calculated as the difference between baseline earnings and counterfactual earnings in 2026.

2.4.10. Earning loss for caregivers

The likelihood of caring for a parent with dementia (model described earlier) and the impact of caregiving on work and the impact of work on earnings were described above. We simulated a baseline earnings measure. In the counterfactual scenario, respondents’ parents did not have and did not acquire dementia between 2010 and 2026. Respondents who were caregivers to parents with dementia in the baseline scenario, did not provide care in the counterfactual, and worked and co‐resided with parents at the same rate as respondents who did not have parents with dementia in the baseline. Earnings were calculated under this counterfactual scenario. The difference between the baseline and counterfactual scenario is the earnings loss for caregivers in 2026.

3. RESULTS OF INTERNAL VALIDATION ASSESSMENT

We conducted an internal validation assessment by comparing USCDM projections to observed outcomes in the HRS. Transition models were estimated using HRS waves from 1998 to 2010, and simulations were run for the period 2012 to 2020. Model estimates of mortality, demographic characteristics, functional status, and disease conditions were compared with corresponding HRS survey data for 2012, 2016, and 2020, and p‐values were reported for tests of statistical differences (Table 4). The composition of the population changed over the time period. In 2010, the sample represented individuals aged 51 and older and because the same cohort was followed forward, the population was aged 61 and older by 2020. Consequently, some later‐year differences reflected both aging and the smaller size of the surviving sample. P‐values reported were based on bootstrapped transition model estimates for USCDM and sample statistics for HRS.

TABLE 4.

Internal validation.

	2012			2016			2020
Parameter	USCDM	HRS	p	USCDM	HRS	p	USCDM	HRS	p
Age (as of July 1)	66.32	66.31	0.960	69.00	68.90	0.659	71.69	71.57	0.574
Black	0.10	0.10	0.984	0.10	0.10	0.706	0.10	0.10	0.374
Hispanic	0.08	0.08	0.693	0.08	0.09	0.550	0.08	0.09	0.344
Male	0.46	0.46	0.105	0.45	0.47	0.000	0.44	0.46	0.000
Any ADLs	0.17	0.16	0.087	0.18	0.17	0.071	0.20	0.17	0.000
Any IADLs	0.14	0.15	0.610	0.15	0.15	0.865	0.16	0.14	0.002
Cancer	0.14	0.14	0.510	0.16	0.16	0.893	0.19	0.19	0.498
Diabetes	0.21	0.21	0.404	0.24	0.25	0.211	0.28	0.28	0.598
Heart disease	0.22	0.22	0.854	0.26	0.25	0.185	0.30	0.27	0.001
Hypertension	0.56	0.56	0.287	0.62	0.60	0.001	0.68	0.63	0.000
Lung disease	0.09	0.09	0.317	0.10	0.11	0.289	0.12	0.12	0.912
Stroke	0.08	0.08	0.879	0.09	0.09	0.133	0.11	0.09	0.000
Dementia	0.05	0.05	0.201	0.05	0.05	0.287	0.06	0.06	0.485
Died	0.06	0.04	0.000	0.06	0.06	0.127	0.07	0.07	0.977

Abbreviations: ADL, activities of daily living; IADL, instrumental activities of daily living; HRS, Health and Retirement Study; USCDM, United States Cost of Dementia Model

The simulated population matched the HRS closely on age, sex, and racial/ethnic composition. Age differences were small, and racial/ethnic shares remained stable across years. The only consistent deviation was a slight under‐representation of males in the USCDM in the later validation years.

Functional status was statistically different only in year 2020 (ADL p‐value 0.002; IADL p‐value 0.002) with quantitative differences of 0.03 and 0.02 percentage points for any ADL, IADL respectively. There were no statistical differences for each disease condition in 2012, and a 0.02 difference for hypertension in 2016 and hypertension, heart disease and stroke in year 2020. There was a 2‐year mortality difference of 0.02 between USCDM and HRS in year 2012 only.

Figures 3 and 4 show data from the HRS for respondents ages 55 and older by sex from 1998 to 2020 and predictions from the USCDM for the population age 55 and older from 2010 to 2040. Across all chronic conditions and functional status outcomes, the USDCM predictions and HRS data exhibit face validity.

FIGURE 3.

Observed and simulated trends in ADL, IADL, dementia, chronic disease, and nursing home prevalence for women aged 55 and older. ADL, activities of daily living; IADL, instrumental activities of daily living.

FIGURE 4.

Observed and simulated trends in ADL, IADL, dementia, chronic disease, nursing home prevalence for men aged 55 and older. ADL, activities of daily living; IADL, instrumental activities of daily living.

4. DISCUSSION

This paper documented the USCDM used to estimate the societal cost of dementia in the United States in 2026, detailing the model structure, data sources, assumptions, valuation methods, and validation. Its aim was to improve transparency, support replication and extension, and guide appropriate interpretation and use of the cost estimates produced by the model. The USCDM, a dynamic microsimulation model, was designed to elucidate and quantify the breadth of costs of living with dementia in the United States. It produced an annual societal cost of dementia.

The USCDM has several strengths. The model built on two established and widely used and well‐validated dynamic microsimulation models. It extended the Future Elderly Model (FEM), a widely used framework for projecting health and economic outcomes among older adults in the United States (e.g. Goldman et al., 2015). ¹⁴ Second, it incorporated and expanded methods developed in the Alzheimer's Disease FEM, which was designed to evaluate the economic value of delaying Alzheimer's disease onset. ¹ The USCDM integrated and advanced these prior approaches to support a comprehensive assessment of the societal cost of dementia.

The model estimated costs associated with living with dementia rather than costs causally attributable to dementia. Because dementia fundamentally alters patterns of care and interacts with multiple comorbidities, costs cannot always be cleanly disentangled from those of co‐occurring conditions. As a result, estimates should be interpreted as societal burden rather than excess costs attributed to dementia. This is appropriate for quantifying the real‐world resource burden on families, health systems, and society by capturing how dementia reshapes medical care, long‐term care, quality of life, and family caregiving in ways that cannot be easily disentangled from co‐occurring conditions. This comprehensive approach enables meaningful comparisons of disease burden across conditions and supports informed targeting of resources.

The USCDM drew on rich, nationally representative data sources, the Health and Retirement Study, Medicare data and other national sources. These data allowed the model to reflect observed patterns of health, care use, and economic outcomes among older adults in the United States and to generate present‐day population‐level estimates relevant for national policy analysis. By using dynamic microsimulation, the USCDM followed individuals over time, allowing costs and outcomes to accrue along heterogeneous disease trajectories. This approach captured variation in dementia onset, survival, care arrangements, and allowed for interactions between health status, caregiving, labor supply, and economic outcomes that cannot be represented in static or prevalence‐based models. This life‐course perspective also enabled the model to quantify the cumulative impacts of dementia on quality of life and earnings for individuals with dementia and their care partners by comparing simulated outcomes to counterfactual trajectories in which dementia does not occur.

A key strength of the model was its explicit representation of unpaid caregiving, including distinctions between spousal and non‐spousal caregivers and valuation of caregiving time using both replacement rate and opportunity of cost methods. By modeling caregiver outcomes and quality‐of‐life effects, the USCDM accounted for substantial economic and welfare impacts that fall outside the healthcare system and had been absent from prior studies. ² , ³ The model was designed to evaluate how changes in prevention, diagnosis, treatment, and care policies may affect future dementia‐related costs. Its flexible structure allows for simulation of alternative scenarios, making it a useful tool for assessing the potential economic implications of interventions and policy reforms.

There were limitations. Like all simulation models, the USCDM relied on the quality and completeness of underlying data. Dementia identification, caregiving intensity, and quality‐of‐life measures are subject to limitations in survey instruments and measurement error, which may affect cost estimates. The omission of the population younger than age 51 sacrifices little generality, since dementia and expenditures on the disease generally occur at older ages. We accounted for care partners of all ages that are reported by respondents who receive care. However, estimates for adult children caregivers are based on HRS respondents at least age 51; thus, we assumed the impacts on earnings or quality of life are quantitatively the same for caregivers of younger ages. Valuing unpaid caregiving time and quality‐of‐life losses required methodological assumptions that were described for transparency but are necessarily uncertain and alternative specifications could be adopted. Broader dimensions of wellbeing losses beyond those captured by care hours, earnings losses and health‐related quality of life of persons living with dementia and care partners are not captured in USCDM. There are spillovers to other family members and friends and losses of well‐being tied to loss of independence and social and personal lives among others that impact well‐being. Future cost estimates will strive to incorporate this.

USCDM relies on survey data from recent HRS respondents. Relationships estimated using those data are assumed to hold into the future. Assumptions about the future of medical progress are modest – USCDM assumes reductions in all‐cause mortality, but no particular pathways for these innovations are imposed or assumed. Advances in disease areas that impact dementia risk and mortality such as heart disease can have unclear impacts on the future cost of dementia, as the risk of dementia might fall while the number of individuals alive to be at risk increases. The structure of the dynamic microsimulation framework is well suited for analyzing these pathways.

To remain tractable and considering data strengths and limitations, the model necessarily simplified aspects of dementia progression, care decision‐making, and institutional placement. These simplifications may not fully capture all clinical or social complexities, particularly at advanced stages of disease. Dynamic microsimulation models are inherently complex and computationally intensive. While the USCDM prioritizes transparency through detailed documentation, the complexity of the model may limit accessibility for some users and constrain the number of scenarios that can be explored. Ongoing work utilizing machine learning to develop a USCDM emulator will attempt to improve accessibility for more users through lowering the computational burden and complexity of working with the simulation. Importantly, the USCDM is constantly being improved with new data sources and methods and validation analysis updated. These updates will be posted on a code repository and in the future, the model will be hosted on the NIA‐funded LINKAGE enclave for broad use by the research community.

5. CONCLUSION

This paper documented the USCDM, a dynamic microsimulation framework designed to estimate the full societal costs of living with dementia in the United States in 2026. By integrating medical and long‐term care spending, unpaid caregiving, earnings losses, and quality‐of‐life impacts for people living with dementia as well as their care partners, the model provided a comprehensive picture of the real‐world burden of the disease. Building on established and well‐validated models and nationally representative data, the model captured heterogeneity in disease trajectories, care arrangements, and economic outcomes over time. Annual cost of dementia in the United States should be interpreted as costs associated with living with dementia in the United States and not dementia‐attributable costs. Estimates were based on available data and informed assumption and therefore carried some uncertainty. Even so, the USCDM provided estimates to inform resource prioritization and targets and offered a flexible and policy‐relevant tool for evaluating how prevention strategies, treatments, and care delivery reforms may influence future dementia‐related costs. This study explicitly documented data sources, measures, methods, and assumptions to promote transparency, reproducibility, comparability with other models and estimates, and appropriate use of dementia cost estimates in research and policy analysis.

CONFLICT OF INTEREST STATEMENT

Dr. Dana Goldman reports grants received in the last 3 years from the American Heart Association, Alexion, Amgen, Biomarin, Blue Cross Blue Shield of Arizona, Blue Cross Blue Shield of Massachusetts, BrightFocus, Bristol Myers Squibb, California Hospital Association, Cedars‐Sinai Health System, Charles Koch Foundation, CommonSpirit, Edwards Lifesciences, Gates Ventures, Genentech, Gilead Sciences, Incyte, Johnson & Johnson, Lilly, National Institute on Aging, National Institute of Diabetes and Digestive and Kidney Diseases, Novartis, Pfizer, RA Capital, and Sarepta Therapeutics; personal fees from Edwards Lifesciences and GRAIL. He is a cofounder of EntityRisk and holds equity in the company. The remaining authors have nothing to disclose. Author disclosures are available in the Supporting Information.

DISCLOSURES

In the past 3 years, Dr. Goldman reports grants from American Heart Association, Alexion, Amgen, Biomarin, Blue Cross Blue Shield of Arizona, Blue Cross Blue Shield of Massachusetts, BrightFocus, Bristol Myers Squibb, California Hospital Association, Cedars‐Sinai Health System, Charles Koch Foundation, CommonSpirit, Edwards Lifesciences, Gates Ventures, Genentech, Gilead Sciences, Incyte, Johnson & Johnson, Lilly, National Institute on Aging, National Institute of Diabetes and Digestive and Kidney Diseases, Novartis, Pfizer, RA Capital, and Sarepta Therapeutics; personal fees from Edwards Lifesciences, and GRAIL. He is a co‐founder of EntityRisk and holds equity in the company. The remaining authors have nothing to disclose.

CONSENT STATEMENT

Consent was not necessary for this study.

Supporting information

Supporting Information