Study design and participants
We used data from four HRS family longitudinal studies, which collected harmonized sociodemographic, economic, health and cognition data for community-dwelling adults from more than 30 countries. The studies included HRS in the United States49; ELSA50; SHARE, which encompasses 28 countries51; and CHARLS52.
The four studies have been designed with similar protocols and frameworks to facilitate cross-national comparisons, which encompass both developed and developing countries. Specifically, the HRS is a nationally representative longitudinal survey of Americans aged 50 years and older, conducted biannually since 1992 with approximately 20,000 respondents per wave49,53. The ELSA includes a nationally representative sample of adults aged 50 years and older in England, with biannual rounds since 2002 with around 10,000 participants per wave50,54. SHARE surveys European adults aged 50 years and older, starting biannually since 2004, initially covering 10 European countries with subsequent expansion to 28 countries (including 27 European countries and Israel) in 2017. The SHARE sample size has exceeded 65,000 since 2013 (ref. 51). Finally, CHARLS is a nationally representative longitudinal survey of about 19,000 Chinese adults aged 45 years and older, initiated in 2011 (refs. 52,55). In all these surveys, participants completed a standardized questionnaire administered either face-to-face or via the internet/telephone. Further details on the sampling and study protocols can be found in the respective sources49,50,51,52,53,54,55.
We constructed variables using harmonized HRS family study data adapted from the RAND HRS and Gateway to Global Aging56,57. These publicly available data sources were harmonized to enhance comparability across the studies. Because only deidentified data were used, our study was deemed exempt from review by the institutional review board at Yale University. Participants in the original studies gave informed consent, and each study was approved by a relevant ethics body49,50,51,52,53,54,55.
Extended Data Fig. 1 shows the sample selection process and final sample sizes. To ensure temporal comparability of measures across surveys, our analysis focused on data from the following sources: HRS (waves 11–14, 2012–2018), ELSA (waves 6–9, 2012–2018), SHARE (waves 5–7, 2013–2017) and CHARLS (waves 2–4, 2013–2018). For each of the four studies, we restricted our samples in each wave to participants aged 50 years and older who reported functional limitations. From this group, we excluded participants who had no valid dementia assessment during the study period. The remaining participants included those who developed dementia (referred to as the dementia sample) and those who never developed dementia (referred to as the non-dementia sample) during the study period. The dementia samples were the primary focus of our analysis, covering the period from 2012 to 2018 and included participants aged 50 years and over, with 1,203 persons (2,717 person-waves) from HRS, 472 persons (1,125 person-waves) from ELSA, 3,041 persons (5,128 person-waves) from SHARE (covering 19 countries; see Supplementary Table 2) and 1,041 persons (2,067 person-waves) from CHARLS. The non-dementia samples were used only for the exploratory analysis.
Dementia assessment
Dementia status was assessed in participants with functional limitations using validated criteria specific to each HRS family study58,59,60,61. For HRS, we employed a well-established algorithm, classifying participants as PLWD if their 27-point cognition summary score was 6 or lower58,59. The 27-point cognition scale comprises three cognitive tests: immediate and delayed word recall tests to measure memory (0–20 points), serial sevens subtraction test for working memory (0–5 points) and counting backwards test for speed of mental processing (0–2 points).
For ELSA, SHARE and CHARLS, an alternative algorithm was used as previously described60,61. Participants with functional limitations were classified as PLWD if their cognition summary scores were 1.5 s.d. below the mean of the population stratified by education levels60,61. As backward counting was not assessed, we used a 25-point cognition scale with the same word recall tests (0–20 points) and serial sevens test (0–5 points) as in HRS. Given the differences in cognitive scales and algorithms between HRS and the other three studies, we performed sensitivity analyses where participants’ cognitive status was evaluated using the same 25-score scale and defined based on the 1.5 s.d. threshold.
For each study, dementia status was assigned in each wave, and participants were classified based on whether they developed dementia during the study period. Proxy assessment of cognition was not considered to ensure comparability across four studies.
Functional limitations and absence of care
Functional limitations were assessed using ADLs and IADLs. The ADLs included six items: dressing, walking across a room, bathing, eating, getting in and out of bed, and toileting, and the IADLs included five items: preparing hot meals, shopping for groceries, making phone calls, taking medications and managing money14,18. ADL and IADL items were assessed similarly in the HRS family studies. The participants were asked whether they had any difficulty with each ADL and IADL because of a health or memory problem or not. This resulted in 11 binary indicators of limitations, one for each ADL and IADL, reflecting different aspects or domains of functional limitations. In this study, we measured the extent of functional limitations by the number of ADL and IADL limitations (sum scores of ADLs and IADLs, range 0–11), ADL limitations (sum scores of ADLs, range 0–6) and IADL limitations (sum scores of ADLs, range 0–5). Definitions and measurements across studies are presented in Supplementary Table 3.
Participants who reported limitations were asked whether they received any care for these activities, with separate questions for ADLs, IADLs and the types of care received (formal versus informal care). To assess the absence of care for functional limitations, we constructed binary variables to indicate whether participants received no care at all for their reported ADL limitations (0/1), IADL limitations (0/1) and either ADL or IADL (0/1) limitations. In other words, the absence of care (for ADLs) meant that the participant did not receive any care for any of their reported (ADL) limitations. Additionally, for each type of functional limitation, we differentiated between formal care and informal care. Therefore, the absence of care was defined respectively for ADLs (including three binary variables: no care, no informal care, no formal care), IADLs (three binary variables) and for ADLs and/or IADLs altogether (three binary variables). Survey questions and their similarities and differences across studies are presented in Supplementary Table 4, and availability of data and variables are provided in Supplementary Table 5.
Statistics and reproducibility
Main analyses
The primary analyses focused on the dementia sample, examining their functional limitations and absence of care. Descriptive statistics of the sociodemographic characteristics, ADL and/or IADL limitations and care received were estimated for each study using pooled person-wave data. Categorical variables were reported as numbers and percentages, whereas continuous/count variables were reported as means and s.d. Missing data for the dementia sample were minimal as shown in Supplementary Table 6.
We used GEE models to estimate wave-specific proportions of absence of care for each study, accounting for within-person correlation over the follow-up period. Each GEE model employed a binomial family with a logit link and specified an exchangeable correlation structure, which fit the data better than other correlation structures, such as first-order autoregressive (AR-1) and unstructured62,63. Supplementary Table 7 provides the estimated correlation (ρ) between pairs of observations for each outcome with more details about correlation structure in the footnotes. Survey weights for the study period (2012–2018) were applied in all analyses to account for sampling design and study attrition.
Interview year was the main independent variable to model trends over time, adjusting for age, sex and the number of ADLs and IADLs. To obtain wave-specific estimates of prevalence for the absence of care, interview years were specified as indicator variables, and predictive margins were used to obtain the adjusted average levels of the absence of care for each year64. The adjusted annual percent change (AAPC) in binary outcomes over time was then modeled with the interview year as a continuous variable. The AAPC was calculated using the formula (ORy − 1) × 100%, where ORy represents the yearly odds ratio65.
We further introduced indicator variables for educational attainment (less versus more educated, stratified by median levels of education) and current-wave living arrangement (living alone versus not living alone), respectively, into the GEE model to examine group differences in the absence of care, while also accounting for covariates and within-person correlation. The statistical significance of each group difference was tested directly in the models and predictive margins were applied subsequently to obtain the adjusted average values for each subgroup.
Sensitivity and additional analyses
To ensure the robustness of our results regarding dementia assessment and sample selection, we conducted two sets of sensitivity analyses for the dementia sample. First, as the cognitive scale and dementia classification algorithm in HRS differed from those in the other three studies, we applied the same 25-point cognitive scales and classification algorithms (1.5 s.d. threshold) used in the other studies to HRS. This ensured consistency in dementia assessment across all studies. Second, to account for potential variations in the timing of dementia onset, we restricted our sample to include only person-waves from the first occurrence of dementia onward.
To further evaluate care burden, we analyzed trends in the number of ADL and/or IADL limitations among the dementia sample. Specifically, we employed GEE models with a Gaussian family and an identity link to estimate wave-specific numbers of ADL and/or IADL limitations over the follow-up period, adjusting for age and sex. The adjusted annual change was estimated using the coefficient of the interview year.
Additionally, to assess differences in the absence of care among people with more severe functional limitations, we repeated our primary analysis using GEE models with a logit link to estimate trends in care absence within the dementia sample with multiple limitations. Specifically, severe functional limitations were defined using the median as a cutoff, identifying people who had difficulty in two or more ADLs (for ADL care provisions), two or more IADLs (for IADL care provisions) and either two or more ADLs or two or more IADLs (for combined ADL and/or IADL care provisions).
Exploratory analyses on non-dementia sample
To determine how the results of the dementia sample differed from the non-dementia sample, we repeated our key analyses for the non-dementia sample. Specifically, we used the same approach to estimate (1) descriptive statistics, (2) trends in the absence of care and (3) trends in the number of functional limitations for the non-dementia sample.
All analyses were conducted using STATA (v.17.0), with two-sided statistical tests and an alpha level of 0.05 for determining statistical significance. Robust standard errors were estimated. The study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
link