Our study utilised a dataset acquired from a prospective cohort study, the Singapore Stroke Study (S3) . Participants were recruited from December 2010 to September 2013 at five public hospitals: Changi General Hospital, Khoo Teck Puat Hospital, Tan Tock Seng Hospital, Singapore General Hospital and National University Hospital . The inclusion criteria for the participants were (1) aged 40 years old and above, (2) Singaporeans or permanent residents, (3) residing in Singapore for the next one year, (4) diagnosed with stroke recently by a clinician and/or verified by CT/MRI brain scan and (5) not having global aphasia . All patients who were eligible were approached for the study.
Our analysis only included participants who were diagnosed with stroke and had a caregiver. A caregiver could be a friend or family member, aged 21 years and older who tended to the patient's needs without financial compensation . We excluded participants who changed caregiver, had no caregiver, or had no information on caregiver status at three-month post-stroke. If the patient changed caregiver at one-year post-stroke, our analysis excluded the caregiver's data at that time point. A total of 399 caregivers were identified from S3 (Fig. 1). After excluding caregivers who were ineligible, 214 caregivers were included in the analysis. At one-year post-stroke, ten patients changed caregivers, and 61 caregivers were lost-to-follow-up.
The S3 collected data from the patient and caregiver at five time points: baseline, three-month, six-month, nine-month and one-year post-stroke . Our analysis focused on three-month and one-year post-stroke, as the variables of interest were available only at these time points. The data was collected via face-to-face interviews at three-month and one-year post-stroke .
Written informed consent was obtained from the patient and caregiver after explaining the study procedure . The study was approved by the SingHealth Centralized Institutional Review Board (2010/724/A) and the National Health Group Domain Specific Review Board (A/10/690) .
The outcome variable was caregivers’ depressive symptoms, measured using the 11-item Center for Epidemiologic Studies Depression (CES-D) instrument  at three-month and one-year post-stroke. It allows respondents to rate various symptoms related to depression using a 3-point Likert scale, with 1 = None/Rarely to 3 = Often. The total score varies between 11 to 33, with a higher score indicating more depressive symptoms. A local study had previously utilised this instrument to measure caregivers’ depressive symptoms .
Variables of interest
Our analysis examined the following caregivers’ psychosocial characteristics: subjective burden, expressive social support and quality of care-relationship between patient and caregiver. Subjective burden refers to the caregivers' adverse psychological and emotional reactions to their caregiving role . Zarit Burden Interview (ZBI) was used to measure this construct. It is a 12-item instrument that allows respondents to rate negatively phrased questions on caregiving from 0 = Never to 4 = Nearly always . The total score varies between 0 to 48, with a higher score indicating a higher subjective burden.
Expressive social support allows caregivers to share their experiences and express their emotions to others . In the process, they can “share sentiments, seek understanding, vent frustration and build up self-esteem” . It was measured using an 8-item scale devised by Pearlin et al. . The responses were captured using a 4-point Likert Scale, with 1 = Strongly Disagree to 4 = Strongly Agree. The composite score ranges from 8 to 32, with a higher score representing better expressive social support. The variable had high internal consistency.
The quality of care-relationship between the patient and caregiver was measured using an instrument devised from the University of Southern California Longitudinal Study of Three-Generation Families . It is a 4-item scale that measure “general closeness, communication, the similarity of views about life and degree of getting along” . Responses were recorded with a 5-point Likert scale, with 1 = Not at all to 4 = Very. The total score ranges between 3 to 12, with a higher score representing a better patient-caregiver relationship. The third question (the similarity of views about life) was removed as local participants had difficulty comprehending it during pilot testing.
The established confounders included for the analysis were: (1) caregiver’s sex, (2) patient’s sex, (3) caregivers’ ethnicity, (4) caregiver's relationship with the patient, (5) patient’s age, (6) caregivers’ chronic conditions, (7) stroke survivor’s depressive symptoms and (8) objective burden. Variable (1)–(4) were significantly associated with caregiver's depressive symptoms in the systematic review by Loh et al. . Variable (5)–(7) were significantly related to caregivers' depressive symptoms in a local study . Several studies related to stroke and other conditions had suggested that caregivers' depressive symptoms were associated with their objective burden [28,29,30]. A stroke-related systematic review  also considered controlling for objective burden as a criterion for a well-designed study. Hence, our study included objective burden as a confounder.
For caregivers’ chronic conditions, caregivers were asked to self-report the presence of 21 health conditions. Following the methods from the local study , eight chronic conditions were selected: arthritis or rheumatism, asthma, cancer or leukemia, cataract, diabetes, heart problems, high blood pressure and kidney diseases. Stroke was not included in the analysis, as it was not captured in S3. The responses were classified into having no health conditions and having one or more health conditions. For stroke survivor’s depressive symptoms, it was measured using 11-item CES-D, which was similar to depressive symptoms of caregiver.
Objective burden includes the time and difficulty faced by caregivers in performing caregiving tasks [14, 31]. It was assessed using Oberst Caregiving Burden Score (OCBS), a 15-item instrument that allows caregivers to rate the time and difficulty for specific caregiving tasks . In S3, the instrument measured only the time aspect. The responses were captured using a 5-point Likert scale, with 1 = none to 5 = A great amount. The total score ranged from 15 to 75, with a higher score indicating that caregivers spend a longer time doing the tasks.
Patients’ and caregivers’ characteristics were presented at baseline. The following variables were included as time-varying variables: caregivers’ depressive symptoms, all variables of interest (subjective caregivers’ burden, expressive social support, quality of care-relationship), caregivers’ chronic conditions, stroke survivors’ depressive symptoms and caregivers’ objective burden. The summary statistics of these time-varying variables were presented at three-month and one-year post-stroke. Continuous variables were presented in either mean (standard deviation (SD)) or median (interquartile range (IQR)), depending on data distribution. Categorical variables were reported in frequencies and percentages.
Caregivers’ CES-D score was a censored variable, as the majority of the respondents had a score of 2 and below (three-month post-stroke: 33.0%, one-month post-stroke: 40.4%, Additional file 1: Fig. 1a, b). Therefore, our analysis used a mixed-effect Tobit regression model with a random intercept, which is suitable for censored outcome variables . A random intercept was included to account for the possible intra-individual correlation in caregivers’ depressive symptoms at three-month and one-year post-stroke. This inclusion helped to estimate the individual intercepts for each caregiver. The mixed-effect regression model is adaptable to incomplete data in repeated measures , which means that caregivers with three-month post-stroke data but missing one-year post-stroke data (n = 71) can be included in the final model (n = 214).
All independent variables (baseline demographics, variables of interest and established control variables) were subjected to bivariate analysis to identify significant variables for the multivariate model. We then removed non-significant variables in the multivariable model using a stepwise approach with the Wald test. The final multivariable model consisted of all significant variables, controlling for the effect of time and established control variables. Partial regression coefficients and 95% confidence interval (CI) were presented for the multivariable model.
All analyses were performed using Stata/IC 16.0 (College Station, Texas), with a two-sided test at a significance level of 5%. The mixed-effect Tobit regression model was implemented using the metobit command . Except for missing one-year data attributed to a change in caregiver or loss to follow-up, missing data were handled via complete case analysis.