We used data from 2002 to 2014 of the publicly available HRS (Core HRS files and RAND 2014 V2 files) . We used data from 2002 onward because they represent the first year HRS changed their approach to identify caregiving support. HRS initially samples from community-dwelling adults, but once a respondent enters the study they are followed until death. Periodically, HRS adds new cohorts to the study so as to maintain a nationally representative sample. Nearly all baseline interviews are conducted face-to-face. Follow up interviews are conducted approximately every 2 years, where half the sample participates face-to-face and the other half participates by telephone. During HRS interviews demographic, economic, family, and medical information are collected on the core respondent and their spouse. Data are also collected on the care respondents receive. If a respondent is unable to participate in a survey, then a proxy respondent is identified.
This secondary analysis of HRS data was approved by the Institutional Review Board of Brown University under protocol 3#1810002244.
ADRD incident cohort
We used the validated Langa-Weir algorithm to identify community-dwelling HRS respondents predicted to have ADRD. The Langa-Weir ADRD algorithm combines the immediate and delayed recall test, serial 7 subtraction test, and backward count from 20 test to generate a wave specific cognitive score to predict if a respondent has normal cognition, cognitive impairment but not dementia, or dementia (i.e., ADRD) [34,35,36]. For respondents with a proxy, the Langa-Weir algorithm predicts respondent cognitive status based on the proxy’s assessment of the respondent’s memory and functional limitations and if the respondent was unable to complete the survey due to cognitive limitations. The Langa-Weir algorithm was validated against ADRD diagnoses from the Aging, Demographics, and Memory Study and was found to correctly classify 76% of self-respondents and 84% of respondents with a proxy .
To identify community-dwelling ADRD incident cases, we identified the first HRS interview a respondent had ADRD between 2002 and 2012 according to the Langa-Weir algorithm. We did not use 2014 data to identify incident cases, as these individuals would automatically be censored following 2014 (the last year of data available in the RAND HRS 2014 V2 files). However, we did follow incident cases identified from 2002 to 2012 up to 2014. Finally, we excluded HRS respondents that had ADRD in one wave but in subsequent waves did not based on the Langa-Weir algorithm.
In addition to the inclusion criteria described above, we excluded persons with ADRD that had data linkage issues across core HRS files and/or that had inconsistent nursing home entry or death dates. We also excluded individuals that had missing data on model covariates (described below). We limited our analysis to up to 8 years post incidence due to the small sample size of individuals with ADRD still in the community at 8 years (n = 30). We also determined if each respondent transitioned to a nursing home, died while in the community, dropped out of the HRS/did not participate in an HRS interview, or remained in the sample/was censored due to the study design.
HRS respondents report if they received help from any individuals (i.e., caregivers) when performing instrumental (IADL; preparing hot meals, shopping for groceries, making telephone calls, and taking medications) and/or basic activities of daily living (ADL; getting across a room, dressing, bathing, eating, getting in/out of bed, and toileting). For each identified caregiver, the respondent reports their relationship to the caregiver and the number of hours and days in a month that caregiver provided assistance. We categorized caregivers as a spouse, adult child, other relative, nonrelative, or paid caregiver. An individual was classified as being a paid caregiver if they were employed by an organization or were a nonrelative that was paid to provide assistance. For each respondent and HRS interview, we calculated the total monthly hours of caregiving received, number of caregiving days in a month caregiving was received (i.e., sum of days of care provided by all caregivers), and number of monthly caregivers. We assumed caregivers could provide at most 16 h of care per day . Our measure of the number of caregiving days could exceed 30 days (e.g., spouse and adult child each provide 20 days of care which is equivalent to 40 care days).
Individual and family characteristics
We obtained the respondent’s age, gender, race (white, African American, other [American Indian, Alaskan Native, Asian, and Pacific Islander]), years of education, number of chronic conditions (0–8; high blood pressure or hypertension, diabetes or high blood sugar, cancer [except skin cancer], lung disease except asthma or emphysema, heart attack/coronary heart disease/angina/congestive heart failure/or other heart problems, stroke, psychiatric problems, or arthritis), whether they were enrolled in Medicaid, whether they had long-term care insurance, and if they had a proxy respondent. The RAND HRS reports whether a respondent has any difficulties (0 = no difficulty; 1 = any difficulty) performing IADLs (preparing hot meals, shopping for groceries, making telephone calls, taking medication; 0–4) and ADLs (getting across a room, dressing, bathing, eating, getting in/out of bed, and toileting; 0–6) . Respondents are asked to only report difficulties that are expected to last more than 3 months. We determined the total number of functional limitations (0–10) a respondent had in each HRS wave by summing the binary indicators across the measures of IADLs/ADLs. Finally, we obtained information on the respondent’s family characteristics including their net worth, marital status, number of sons, number of daughters, number of married children, number of living siblings, number of grandchildren, and number of great grandchildren.
We evaluated the characteristics of the person with ADRD, their family, and the amount of caregiving received in the community approximately every 2 years (mean time between interviews was 1.95 years (SD = 0.35)) from incidence up to 8-years post incidence. At each time post incidence, we also evaluated the characteristics of those who transitioned to a nursing home or died in the community.
In preliminary analyses, we observed that the measures of caregiving received were associated with transitions to nursing homes and mortality. Therefore, to estimate the association between individual and family-level characteristics of the person with ADRD and receiving care in the community we estimated a series of joint mixed effects and survival models . We estimated separate joint models for each measure of caregiving outcome (hours of caregiving, days of caregiving, and number of caregivers). The joint modeling framework consists of two parts that share parameters and the models are estimated with a joint likelihood function. In the first part, we estimated a random intercept and slope mixed effects model where the outcome represents the measure of caregiving received (hours of caregiving received, caregiving days, and number of caregivers). In the second part, we estimated a Weibull survival model in which the outcome represented nursing home placement or mortality in the community (whichever occurred first). Withdrawal from the HRS, missing an HRS interview, and being alive and in the community in 2014 or at 8-years post incidence were treated as censoring events. We modeled the association between the mixed effects and Weibull survival model using the current value parameterization method .