Study population
This study used mortality follow-up data from a survey carried out between 1998 and 1999 in a representative sample of people 65 years or older living in care homes in Madrid, Spain. The sample was assembled by means of a stratified cluster sampling, by selecting 25 public/subsidized and 30 private facilities with probability proportional to their size, and then randomly sampling 10 men and 10 women from each selected public/subsidized facility, and 5 men and 5 women in private institutions. Of the 800 sample residents, 85 subjects refused to participate (response rate 89%) and 39 of these were randomly replaced with residents of the same facility and sex, for a total of 754 participants in the baseline survey. Residents in public/subsidized facilities and men were oversampled; accordingly, sampling weights were computed for study subjects as the inverse of their probability of selection. Subjects were eligible for the present study if they were residents of nursing homes in Madrid, Spain, aged ≥65 years, and did not suffer from severe physical or cognitive impairments at baseline (severe or total functional dependency, severe cognitive impairment, physician’s diagnosis of dementia, or behavioral problems).
Baseline data collection
Trained geriatricians and residents in geriatrics interviewed the residents, their main caregivers, the facility staff, and the facility physicians to gather information on sociodemographic data, facility characteristics, SE within the facility, external contact with family and friends, morbidity, functional dependency, cognitive status, and behavioral problems.
By interviewing residents (or next of kin) we obtained information about their age, sex, education, and marital status. By interviewing the facility staff, we obtained information about facility size, caregivers assigned to residents, and length of stay in the facility.
Residents’ SE levels within the nursing home were determined according to each resident’s degree of interaction with other residents, and to their level of active participation in the activities of the facility. To this purpose, study subjects (86%) or their main caregivers (14%) were asked the following question: “To what extent do you/does the resident interact with other residents in the institution: 1. A lot; actively engages in the facility activities; 2. Sufficiently, normal; 3. Hardly; 4. Not at all”. Participants were classified into three levels of SE according to the option selected: high (option 1), moderate SE (option 2), and low/null (options 3 and 4). Information about frequency of external contacts was also obtained from the study subjects or their main caregivers, and classified as monthly or less, weekly, or daily.
We ascertained chronic diseases by interviewing facility physicians (or nurses for 8% of residents) with access to medical histories. These included chronic obstructive pulmonary disease, ischemic heart disease, heart failure, arrhythmias, peripheral arterial disease, stroke, high blood pressure, diabetes, anemia, Alzheimer’s disease, other dementias, Parkinson’s disease, epilepsy, depression, anxiety, arthritis, and cancer. Dementia was defined as a physician’s diagnosis of Alzheimer’s disease or other dementias. The number of these conditions, not including dementia, was registered and categorized into 0–1, 2–3, and ≥ 4 diseases.
Functional dependency in performing activities of daily living was evaluated by interviewing residents or their main caregivers using a Barthel index version [
12]. According to this version [
12], residents were categorized as independent (100 points), mild dependency (91–99 points), moderate dependency (61–90 points), and severe or total dependency (0–60 points).
Cognitive status was evaluated with the Pfeiffer’s Short Portable Mental Status Questionnaire (0–10 errors) [
13], which was adapted to the institutional setting, and the Minimum Data Set Cognition Scale (0–10 points) [
14], which obtained an assessment from the main caregivers based on particular Minimum Data Set questions. Severe cognitive impairment was defined as 8 or more education-adjusted errors in the Short Portable Mental Status Questionnaire or 9 or more points on the Minimum Data Set Cognition Scale. Residents with behavioral problems related to verbal, physical abuse, or inappropriate/disruptive behavior during the previous week were identified through the corresponding Minimum Data Set questions answered by caregivers.
Statistical analysis
The cumulative all-cause mortality curves for each baseline level of SE within the nursing home (low/null, moderate, or high) were standardized to the weighted distribution of baseline confounders in the overall institutionalized population by using inverse probability weighting. We first fitted a sampling-weighted polytomous logistic model to estimate each resident’s probability of being in their own level of SE given the observed confounders. Standardization weights were computed as the inverse of these conditional probabilities, and were further rescaled by the sampling-weighted proportions in each level of SE to reduce variability of weights and avoid influential observations with extreme weights [
16]. We then assigned combined weights to residents as the product of the sampling weights and the standardization weights to correct for both selection bias and confounding [
17].
Three increasingly comprehensive sets of baseline confounders were included in the polytomous logistic model for baseline SE. The first model included age (65–74, 75–79, 80–84, 85–89, or ≥ 90 years), sex (women or men), education (less than primary; primary; or secondary or more), and marital status (married, single, or widowed/divorced). The second model additionally controlled for facility ownership (public/subsidized or private), facility size (< 100, 100–299, or ≥ 300 beds), length of stay (0–1, 2–4, or ≥ 5 years), assigned caregiver (yes or no), and frequency of external visits (monthly or less, weekly, or daily). The third model further adjusted for baseline multimorbidity and disability, including number of diseases (0–1, 2–3, or ≥ 4) and functional dependency (no, mild, or moderate). The mean (range) combined weights, taking into account both sampling and standardization based on these models, were 0.99 (0.18–3.25), 0.99 (0.14–4.12), and 0.98 (0.13–3.77) for the first, second and third model, respectively (Supplementary Figure
1). This weighting procedure provided reliable standardization, as can be appreciated by the fact that the weighted distributions of confounders were almost equal across the different levels of SE, and also narrowly matched their sampling-weighted distributions in the entire institutionalized population (Supplementary Table
1).
For estimating mortality risks, we used Kaplan-Meier methods and flexible parametric survival models using splines [
18], weighted by combined weights and stratified by level of SE to obtain nonparametric and smooth estimates of the cumulative mortality curves that would have been observed in the overall institutionalized population had every resident been in each level of SE [
19]. For models based on splines, stratum-specific log cumulative hazards were modelled as distinct natural cubic splines of log time with two internal knots at the 33th and 67th percentiles [
18]. We used these weighted spline-based survival models to compute standardized differences in cumulative mortality at 2, 5, and 10 years of follow-up, as well as standardized differences in median survival times, for moderate and high SE compared with low/null SE. The 95% confidence intervals (CIs) were derived by applying delta methods on robust standard errors of spline coefficients.
We assessed potential modifications in risk differences among relevant subgroups of residents based on baseline age, sex, facility ownership, facility size, and baseline functional dependency by using weighted spline-based survival models stratified by baseline level of SE and participant subgroup. Weights for subgroup analyses were calculated as the product of sampling weights and subgroup-specific standardization weights, which allowed to standardize cumulative mortality curves for each SE level and resident subgroup to the weighted distribution of confounders in the entire resident subgroup [
16]. We estimated standardized differences in 5-year cumulative mortality and their 95% CIs within each subgroup, and tested for heterogeneity across subgroups by using joint Wald tests. Statistical analyses were run with the
stpm command in Stata, version 14 (Stata Corp., College Station, Texas) and graphics were produced in R, version 4 (R Foundation for Statistical Computing, Vienna, Austria).