Study observations
The CHARLS was a nationally representative longitudinal survey targeting Chinese community-dwelling individuals aged 45 years and older along with their spouses, and used a multistage stratified probability-proportionate-to-size sampling to cover 28 provinces, 150 countries/districts, and 450 villages/urban communities across China [14]. The CHARLS collected information on various demographic characteristics, physical function, chronic disease, family structure, work, socioeconomic status, retirement and pension, health care and insurance, income and consumption. We used data from four waves (i.e., 2011/2012, 2013, 2015, and 2018) of the CHARLS. Details of the CHARLS are provided elsewhere [14].
Participants can participate in one or more waves of CHARLS, and their cohabitation and cognitive function may vary across waves. Therefore, the study unit was defined as an observation rather than a participant. As shown in Fig. 1, we first excluded 1,950 participants due to missing covariates data, from a total of 25,370 participants in any of the 4 waves of CHARLS. Among the 48,126 observations from these remaining 21,405 participants, we excluded 1,984 observations of participants living alone with normal cognition and 7,801 observations of participants living with others and with cognitive impairment. Thus, we identified 909 observations of participants living alone with cognitive impairment (cases) and 37,432 observations of participants living with others and with normal cognition (comparators).
Measurement of cognitive impairment
Based on similar concepts in Health and Retirement Study (HRS), the cognitive function in CHARLS includes episodic memory and executive function. Episodic memory was measured by the immediate and delayed recall of word. 10 unrelated Chinese words were read to each participant, and his/her memory ability was evaluated by adding up the number of correct words recalled immediately (immediate word recall scores) and 4 minutes later (delayed word recall scores). We calculated episodic memory as the average scores of immediate and delayed word recall scores [15], with a range of 0 to 10. Executive function was based on the components of the Telephone Interview of Cognitive Status (TICS) and figure drawing [16]. Components in TICS included today’s date (month, day year and seasons), the day of the week and serially 7 subtracting from 100 (for 5 times). In figure drawing, participants were asked to redraw a picture from one painting. We calculated executive function as the total score from TICS and figure drawing, ranging from 0 to 11. The cognitive function score was the sum of episodic memory and executive function, with a higher score indicating a better cognitive function (range: 0–21) [17]. As done in literature [18], we defined the observation as having cognitive impairment if the summary score was less than 6; otherwise, the observation was defined as having normal cognition.
Measurement of living alone
The CHARLS used the household roster file to collect the number of residents living in a household, including all reported members living in the household, with the exception of the household respondents and their spouse (as identified by the household respondents). Living alone was defined if the number of residents living in household was one [19]; and living with others was defined if the number of residents living in household was greater than one.
Measurement of catastrophic health expenditure
Self-reported information on money that participants paid OOP for their last month’s outpatient visits and last year’s inpatient visits was collected in CHARLS. The spouses of all participants were collected for the same information. Each participant’s annual OOP cost on outpatient care was calculated as the result of multiplying last month’s cost by 12 [6].
A household’s capacity to pay was defined as the total cost of the household’s consumption minus the food-based household cost [6]. A household’s OOP cost on healthcare was defined as the sum of annual OOP cost on inpatient and outpatient healthcare of both the participant and his/her spouse [6]. A household was defined as incurring catastrophic health expenditure if OOP cost on healthcare was ≥ 40% of a household’s capacity to pay [20, 21]. In particular, for those participants who did not have spouses living together in a household, we considered their spouses’ annual OOP cost on healthcare was 0. A binary variable was defined to indicate whether there was catastrophic health expenditure in participant’s household, which has been widely used in previous studies based on CHARLS [6, 22, 23].
Covariates
We considered a series of covariates which may be associated with living alone, cognitive impairment [24–30] or catastrophic health expenditure [31]. These covariates included age, sex (male vs female), and residence status (rural vs others (“city/town”, “combination zone between urban and rural areas” and “special area”), marital status (currently married vs others including “separated”, “divorced”, “widowed”, and “never married”), education (no schooling vs primary school or more), alcohol consumption (non-drinker vs drinker), smoking status (non-smoker, ever smoker and current smoker), and disease counts. Disease counts were calculated based on the number of self-reported diseases diagnosed by doctor, including hypertension, cancer, diabetes, lung disease, stroke, heart disease, arthritis, kidney disease, asthma, and digestive disease.
Statistical analyses
In descriptive analysis, mean ± SD was used for continuous variables and numbers and percentages for categorical variables, unless otherwise specified.
To address data imbalance and confounding factors between cases and comparators, propensity score matching was used. A propensity score is the conditional probability of an exposure for a set of covariates [32], which was estimated by a multivariable logistic regression model [33]. The dependent variable was living alone with cognitive impairment, and independent variables age, sex, education, marital status, residence status, smoking status, alcohol consumption, and disease counts. A 1:2 matching protocol was used for matching without replacement (greedy-matching algorithm), and the caliper width was equal to 0.2 of the standard deviation of the logit of the propensity score. We estimated the standardized mean differences (SMDs) for all the covariates before and after matching to assess prematch imbalance and postmatch balance. For a given covariate, a SMD of < 10.0% represented a relatively small imbalance [34].
To investigate whether adults living alone with cognitive impairment had a higher percentage of catastrophic health expenditure than those living with others and with normal cognition, we compared the cases and the matched comparators. We first compared distributions of outcomes (i.e., catastrophic health expenditure) of observations between the cases and the matched comparators using Mann-Whitney U tests for continuous variables and Chi-square test for categorical variables. Second, generalized estimating equation models [35] were used to estimate the odds ratio (OR) and corresponding 95% CIs of catastrophic health expenditure for the cases relative to the matched comparators. To account for the correlation between observations from the same participant we used logit link function and autoregressive correlation matrix. We considered three models. Model 1 was a crude model; Model 2 adjusted for age and sex; and Model 3 additionally adjust for education, marital status, residence status, alcohol consumption, smoking status, and disease counts.
We performed two additional analyses to test the robustness of our findings. First, we changed the cut-off value of defining cognitive impairment as the summary score at least one standard deviation (SD) below age-appropriate norms, then re-examined the association with the same models. Second, to further examine whether the association between living alone with cognitive impairment and catastrophic health expenditure differed by some social-demographic characteristics, we conducted the subgroup analyses by age, sex, education, marital status, residence status, alcohol consumption, smoking status, and disease counts.
R version 4.1 and SAS version 9.4 (SAS Institute, Cary, NC) were used to perform all statistical analyses. Statistical significance was defined as a P value less than 0.05 (two-tailed).