Method

Participants

Participants were informant proxies for members of the Puerto Rican Elder: Health Conditions (PREHCO) Study (Palloni et al., 2005) who were part of the 2021/2022 data collection (Wave 3; see Figure 1 for the participant flowchart). The PREHCO study is a representative longitudinal survey of older adults in Puerto Rico, which originated in 2002. In Wave 3, PREHCO participants were aged 78 years and older, and an ancillary PREHCO data collection was developed for caregivers of PREHCO participants who also served as proxies for the main study. Initially, 194 proxies were identified, of whom 132 were a caregiver for the participant and completed the ancillary interview (68%).

Figure 1: Participant flow diagram for the PREHCO dementia caregiver study.

Since this study focused specifically on caregivers of people living with dementia, the 132 caregivers who completed the interviews were further classified as dementia caregivers or non-dementia caregivers based on informant reports. To identify caregivers of PREHCO participants with dementia, we used the Health and Retirement Study (Crimmins et al., 2011) version of the Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE; Jorm, 1994), which asked caregivers about changes over the past two years in memory and everyday function of the older adults for whom they cared. Based on the criteria established by Mejía-Arango et al. (2020), a score of 3.4 or higher indicated that the older adult was likely experiencing dementia, thereby classifying the caregiver as a dementia caregiver. In all, 108 caregivers were identified as caring for people living with dementia. Of these, 101 caregivers with complete information on all key variables used in the current study were included in final analyses. The internal review boards at University of Puerto Rico and University of Alabama at Birmingham approved the original data collection. The internal review board at Arizona State University approved the analyses of de-identified data conducted by the study authors.

Measures

Dependent variable

Depressive symptoms were measured using the Geriatric Depression Scale (GDS), the 15-item version, with binary responses (yes/no) (Sheikh & Yesavage, 1986). To ensure that higher scores indicate greater depressive symptomatology, four of the 15 items were reverse-coded, and possible total scores range from 0 to 15. The GDS has been widely used in previous studies to assess depressive symptoms in dementia caregivers (Gaugler et al., 2018), higher scores reflect higher depressive symptoms. In this study, the 15-item GDS demonstrated acceptable internal consistency (Cronbach’s α = 0.74), and its psychometric properties have been validated in U.S. Latino informal caregivers (Perales-Puchalt et al., 2025).

Independent variable

Caregiver burden was measured using the 6-item Zarit Burden Interview (ZBI). Originally introduced as a 29-item scale (Zarit et al., 1985), the ZBI was later revised to a 22-item version, which became widely used. To facilitate quicker administration, various shorter versions (e.g., ZBI-12, ZBI-8, ZBI-7, ZBI-6, and ZBI-4) were developed. Among these, the ZBI-6, used in this study, has been shown to have the minimal number of items while maintaining diagnostic utility close to the 22-item version (Yu et al., 2019). Each item was measured on a five-point scale from 0 (never) to 4 (very often). Caregiver burden was calculated as the total score across six items, ranging from 0 to 24, with higher scores indicating greater burden. The scale demonstrated good internal consistency in the present study (Cronbach’s α = 0.84). The ZBI-6 has likewise been validated with good psychometric properties among U.S. Latino informal caregivers (Perales-Puchalt et al., 2025).

Moderator variable

Behavioral and Psychological Symptoms of Dementia were measured using the Neuropsychiatric Inventory Questionnaire (NPI-Q) (Kaufer et al., 2000). The NPI-Q assesses BPSD across 12 symptom areas—delusions, hallucinations, depression, anxiety, apathy, irritability, elation, disinhibition, agitation, aberrant motor behavior, appetite and eating disturbance, and nighttime behavioral disturbances, with responses collected based on caregivers’ reports. Caregivers responded to each domain question with “yes” or “no” to indicate the presence or absence of symptoms. If symptoms are present, their severity is rated as 1 = mild, 2 = moderate, or 3 = ­severe. In this study, each domain was coded as a single variable for factor analysis of BPSD. A response of “no” was coded as 0, while a response of “yes” was coded based on severity, ranging from 1 (mild) to 3 (severe). Therefore, each of the 12 domains has a possible score range of 0–3.

Other variables

Based on prior research indicating their influence on caregiver appraisals and mental health, participants’ demographic and health-related characteristics and an indicator of caregiving demand were included as covariates (Perkins et al., 2013), including age (years), sex (0 = male, 1 = female), educational attainment (0 = less than high school, 1 = high school, 2 = associate/technical degree, 3 = bachelor’s degree or higher), marital status (0 = without a partner, 1 = married/partnered), self-rated health (0 = poor/fair, 1 = good, 2 = very good/excellent), and weekly caregiving hours (0 = 35 hr or less, 1 = more than 35 hr up to 100 hr, 2 = more than 100 hr). Perceived social support was included to test whether this variable may partially explain the association between caregiver burden and depressive symptoms. It was assessed using the Lubben Social Network Scale (LSNS-6; Lubben et al., 2006) which measures caregivers’ perceived support from family and friends. Three questions were asked separately for each group: (1) How many people do you see or hear from at least once a month? (2) How many people do you feel at ease with to discuss private matters? (3) How many people do you feel close to and could call on for help? Responses to each question were recorded using six options: 0 (none), 1 (one), 2 (two), 3 (three or four), 4 (five to eight), and 5 (nine or more). Higher total scores indicate greater levels of perceived social support, with a range of 0–30, and the scale demonstrated acceptable internal consistency (Cronbach’s α = 0.77).

Analyses

Descriptive statistics for the participants were used to summarize the characteristics of variables. The 12 BPSD domains measured by the NPI-Q were examined using exploratory factor analysis (EFA) followed by confirmatory factor analysis (CFA). To enhance estimation accuracy, given that the BPSD were measured as ordinal variables, a polychoric correlation matrix was used for both factor analyses. Before conducting EFA, suitability of data was assessed using the Kaiser-Meyer-Olkin (KMO) test and Bartlett’s test of sphericity, with a KMO value of 0.60 or higher and a significant Bartlett’s test (p < .05) indicating suitability for factor analysis. The number of factors was determined using the criterion of eigenvalues greater than 1.0 and by examining the scree plot for a sharp drop in eigenvalues between factors. Items were subsequently assigned to factors using a minimum loading threshold of 0.40. CFA was performed by linking each BPSD item derived from EFA, employing the weighted least square mean and variance-adjusted (WLSMV) estimator, which is appropriate for polychoric correlations (Finney & DiStefano, 2006). Model fit was assessed using comparative fit indices (CFI), Tucker–Lewis index (TLI), and RMSEA, with CFI and TLI values greater than 0.95 and an RMSEA value less than 0.06 indicating good model fit (Hu & Bentler, 1999). Since CFA was conducted using a polychoric correlation matrix, scaled goodness-of-fit indices were used to evaluate model. Finally, the factor scores obtained from CFA were standardized and used in the main analyses.

Multivariable ordinary least square (OLS) regression analyses were conducted to test the research hypotheses. Age, sex, educational attainment, marital status, self-rated health, weekly caregiving hours, and perceived social support were included as covariates sequentially. In Model 1, age, sex, education, marital status, and self-rated health were controlled for. In Model 2, hours of caregiving per week were additionally adjusted. Finally, we added perceived social support in Model 3 to test whether it may explain away any association between caregiver burden and depressive symptoms as per our third hypothesis.

Finally, we examined the interaction effects between each BPSD-based factor and caregiver burden separately. Standardized coefficients were reported to increase interpretability and allow comparisons of the magnitude of associations across models. The Johnson-Neyman technique (Johnson & Fay, 1950) was used to determine the specific range of observed BPSD factor scores within which the association between caregiver burden and depressive symptoms was statistically significant.

All data preprocessing and statistical analyses were performed using R version 4.4.1. Moderation analyses were conducted using the interactions package (version 1.2.0). The first author used ChatGPT to review the completed manuscript for English translation and readability. The content was subsequently reviewed and further edited by all authors.