Data Source and Participants
This study analyzed data from a longitudinal, observational study conducted to explore the care-specific emotional expressions of persons living with dementia. Participants were eligible for the study if they were 65 years or older, diagnosed with dementia based on criteria from the Diagnostic and Statistical Manual of Mental Disorders, fourth edition, and had a Korean Mini Mental State Examination (K-MMSE) score lower than 24. For each participant, nine videos were taken at 0, 3, and 6 months for each of the three specific care situations including mealtime, personal care, and social activity. Since three participants dropped out of the study at three or six months, 30 participants with 258 videos were produced. Of these, this present study analyzed 86 mealtime videos from 30 participants.
Variables and Measures
Person-Level Data
Person-level data, collected only at baseline from participants’ medical chart included demographic data such as age, sex, and education level, and medical information such as comorbidities and medications. Comorbidities were categorized according to the Cumulative Illness Rating Scale-Geriatric (CIRS-G) [14]. This tool is organized into 14 categories, and for diseases present in each category, a minimum of zero and a maximum of four points can be given according to its severity. The illness severity is calculated with the total score (0-56) divided by the number of corresponding categories (0-14). All medications taken by participants were classified as cardiovascular, diabetes, dementia, psychiatric, and other medications.
Observation-Level Data
Observation-level data, collected at 0, 3, and 6 months from the parent study included participants’ behavioral symptoms, caregiving approaches, cognitive status, depression, and function using measures described below.
Participants’ behavioral symptoms were evaluated using the Korean version of Cohen-Mansfield Agitation Inventory (CMAI-K) [15, 16]. This tool is composed of 29 items and organized into three categories: aggressive behavior (e.g., hitting, grabbing, verbal aggression, and so on), physically nonaggressive behavior (e.g., repetitious mannerisms, general restlessness, and so on), and verbally agitated behavior (e.g., complaining, negativism, repetitious sentences, and so on). A higher score means more behavioral symptoms in the original tool [17].
To measure caregiving approaches during mealtime, we used a behavioral coding scheme, which contained items from the Person-Centered Behavior Inventory (PCBI) and Task-Centered Behavior Inventory (TCBI) [18, 19]. This coding scheme measured mealtime interaction between persons living with dementia and caregivers, and comprised verbal and nonverbal behaviors for each category of PCBI and TCBI. The verbal PCBI includes ten behaviors: greeting, asking the resident for help/cooperation, giving choices, assessing comfort, providing orientation, showing approval/interest/empathy, positive voice quality, and back-channel response. The nonverbal PCBI includes seven behaviors: resident-directed eye gaze, positive gestures, appropriate use of affectionate touch, assessing comfort nonverbally, adjusting to the resident’s pace, proximity, and positive facial expressions. The verbal TCBI includes four behaviors: verbal controlling (i.e., interfering/directing tone or elderspeak), interrupting, changing topics, and controlling voice quality. The nonverbal TCBI includes four behaviors: ignoring, physically controlling, inappropriate touch, and outpacing. In addition to the PCBI/TCBI, no interaction codes (i.e., no verbal response and no nonverbal response) were added to the scheme to cover the entire video because there were parts of the video where caregivers did not display any verbal/nonverbal behaviors.
Cognitive status was measured by the K-MMSE and the Korean version of Clinical Dementia Rating (K-CDR). The K-MMSE is a 30-item scale with a total of 30 points; a lower score indicates more impairment in cognition. Untestable cases were given -1 point in the K-MMSE [20]. The K-CDR is a 6-item scale with a total of 30 points, and its higher score denotes more severe dementia [21]. Depression was measured using the Korean version of Cornell Scale for Depression in Dementia (K-CSDD), which consists of 19 items with a total of 38 points, and a higher score implies a more depressive state [22]. Function was rated using the Korean version of Activities of Daily Living (K-ADL), which consists of seven items, with a higher score indicating more dependency [23, 24].
Procedure
After obtaining Institutional Review Board (IRB) approval, we extracted 86 mealtime videos (i.e., eating and feeding) from the parent study, which included 258 videos about three care situations (mealtime, personal care, and social activity). Since the videos contained images and voices of participants and caregivers that would allow them to be recognized, two research assistants (RAs) who managed and coded the videos were particularly well-trained on ethical issues such as confidentiality before starting the study. To ensure intra-rater/inter-rater reliability, we initially sampled 10% of the videos, and two coders coded the same video and compared results. After inter-rater reliability as measured by the kappa statistic reached 0.8, the coding process was started. To ensure reliability between coders throughout the coding process, we conducted an additional reliability check after half the videos were coded.
We used the Noldus Observer® XT software for coding, using the items from the PCBI/TCBI and CMAI as codes. Among the PCBI and TCBI, the verbal behaviors were coded using instantaneous sampling and the nonverbal behavior were coded using continuous sampling. Instantaneous sampling counts how many times each verbal behavior appeared in the entire video (frequency) as the coder assigns the code that corresponded to specific caregiver behaviors every ten seconds. With continuous sampling, the coder assigns the code of nonverbal behaviors of interest whenever they began and finished so that the total time for each behavior could be calculated in seconds (duration) [25]. When caregivers showed no behaviors, the coder put no verbal response or no nonverbal response to provide mutually exclusive and exhaustive codes. Consistent with the PCBI/TCBI coding scheme, no interaction codes were calculated as a frequency for no verbal response and duration for no nonverbal response. The participant’s behavioral symptoms were coded using the CMAI, which was also obtained through continuous sampling. We summed the total time for each behavioral symptom observed and categorized these behaviors into the three categories of the CMAI: aggressive behavior, physically nonaggressive behavior, and verbally agitated behavior. Throughout the whole process, when the coder could not decide which category a behavior belonged to, it was determined after a discussion with another researcher at the weekly meetings.
Statistical analysis
For statistical analysis of the frequency data, we counted the total number of times each behavior occurred in each video and grouped into person- or task-centered verbal behaviors. Then, the total frequency of person- or task-centered verbal behaviors were divided by the total time of each video (frequency per minute) since the total time of each video varied according to participants’ eating pace. Similarly, we added the total seconds of each person-/task-centered nonverbal behavior of the caregivers and behavioral symptoms of the participant and divided the total time by the total time of each video (duration per minute) to account for the difference in video length; therefore, duration per minute referred to total seconds in which a specific action appeared in one minute.
Statistical analysis was conducted using the STATA 16.0 software (StataCorp, College Station, Texas, USA) and consisted of the following steps: 1) descriptive analysis to understand participants’ characteristics, 2) a mixed-effect model to examine whether caregiving approaches were associated with participants’ behavioral symptoms, and 3) a further analysis to determine which particular behavior among significant PCBI/TCBI, were associated with behavioral symptoms. The mixed-effect model was chosen because the data were repeated measures and therefore nested within participants. Both time points and facility were considered in the model. We controlled for the time variable (i.e., time points) to reflect change over time in behavioral symptoms; the facility was included to account for correlation for responses from participants of the same facility. Among various mixed-effect models, we used multilevel mixed-effects tobit regression fixing the lower limit at the minimum value of the dependent variable because behavioral symptom measures were continuous variables and left-censored with many zero values [26].