This study used the data of the “National Evaluation of the EACHD program” project, which was a 16-month time-series study implemented between September 2007 and February 2009 . All fieldwork methods were approved by La Trobe University Human Ethics Committee (07–084).
Population of the original project and data collection
Participant recruitment was undertaken between September 2007 and 2008 and data collection continued until February 2009. According to the government guideline , participant eligibility criteria were:
People aged 70 and over (or 50 and over if indigenous) living in the community or retirement villages.
Having a higher level of residential care needs and the following characteristics: experiencing Behavioural and Psychological Symptoms of Dementia (BPSD); preferring to receive EACHD program services; and being able to live at home with the support and services provided by the program.
Possibly having ADL and IADL disabilities and a higher level of care needs associated with their behaviours of concern, and facing the risk of unavoidable permanent nursing home admission.
However, it should be noted that at the initial stage of the EACHD program, not all 354 individuals enrolled by the original research project were 70 years old or over. Specifically, 27 clients were aged between 58 and 69, and 19 clients were aged between 58 and 64 (strictly speaking not “older people”). In addition, no indigenous people were enrolled even though they were included in the target population of the EACHD program.
Clients and carers’ information was collected from their case managers at baseline (the study entry), and every three months thereafter (up to four assessment points in total). Specifically, clients’ socio-demographic characteristics were collected at baseline. Clients’ medical diagnoses, physical and cognitive status, behavioural problems, services use, carers’ stress due to client BPSD etc. during the past three months were surveyed at baseline and each assessment point. A discharge survey was conducted if a client was discharged within the 16-month study period due to the following reasons: death, permanent nursing home admission, permanent nursing home admission and death, hospital admission, hospital admission and death, hospital admission and nursing home admission, and other unknown reasons.
The number of surveys per client ranged from one (only baseline survey) to five (baseline survey plus four follow-up surveys, or baseline survey plus three follow-up surveys and the discharge survey), depending on the enrolment time - later enrolees had fewer surveys.
Aiming to compile a national data set of at least 300 clients at the initial stage of the EACHD program, the original project stopped recruitment at 354 clients due to time constraints .
Population of our study
We only used 284 participants’ data. The other 70 participants’ data were inappropriate for Cox regression analysis because 65 clients only had the baseline survey and five clients had baseline and discharge surveys occurring at the same dates. The 70 participants were lost to follow up according to the original project. Excluding these participants should not lead to obviously biased results because their baseline variables were similar to those of the 284 participants.
Measures of our study
Outcome variables included (inpatient) death or hospital admission (on either condition clients would be discharged) during six, nine, and 16 months after enrolment. Emergency department (ED) visits and outpatient hospital admissions were not included because clients experiencing these events remained using EACHD program services.
In our study, death and hospital admission were combined as one endpoint—death or hospital admission. One consideration was that time to hospital admission and time to death were likely to be dependent because poor health status potentially increases the risk of both hospital admission and death . In addition in health and medical research, it is not uncommon for researchers to combine adverse health events, such as the permanent nursing home admission and death, and hospital admission and death [3, 32–34]. Another consideration was that small numbers of clients experiencing hospital admission and death made it inappropriate to conduct analyses for the two outcomes separately (see details in Results section).
We also examined risk factors for time to death as among those clients experiencing death or hospital admission most were death cases. As described in the Background, we treated this as a secondary study purpose because there has been similar research on this topic. In addition, we were aware that the findings might be limited by inadequate sample size (see Sample size section below).
Time to death or hospital admission (and death)
Sixteen-month study period
Time to death or hospital admission was estimated as the number of days from the study entry (the date of completing the baseline survey) to the date of death or hospital admission. Clients who had not yet died or moved to hospital were censored at the date of the last survey (last follow-up survey, or discharge survey due to other reasons but not death or hospital admission).
Six and nine-month study periods
Time to death or hospital admission was counted as the number of days from the study entry to the date of death or hospital admission. Clients not experiencing death or hospital admission were censored at the date of the last survey if it occurred before the end of the study period. Or they were censored at the end of the study periods (at six months and nine months) if the last survey occurred after the end of the study periods.
Based on previous studies [6, 14, 18, 20], and available information of the original project, our study included the following baseline variables for Cox regression analysis:
Socio-demographic characteristics: age, gender, birthplaces, income sources, first language, living arrangements, carer relationships, and carer status.
Use (0 = no; 1 = yes) of services: GP visits, ED visits, outpatient visits, inpatient hospital admissions, home nursing care, dementia specialist care, and community care services (including allied health, personal care, domestic care, information services, social support services and respite services).
Use of case management time: measured by number of hours.
Severity of medical conditions: According to the Charlson index of co-morbidity , different diseases of clients were weighted differently. For example, cerebrovascular diseases, diabetes, congestive heart failure etc. scored 1; renal impairment, tumours etc. scored 2; and metatastatic solid tumours scored 6. Based on the total scores of clients’ diseases, clients’ medical conditions were classified into no, mild (1 or 2 scores), moderate (3 or 4 scores) and severe conditions (over 4 scores).
Other health conditions (0 = no; 1 = yes): depression and falls
Global deterioration scale (GDS) score: ranging from very severe problem (coded as 1) to none (coded as 7). Higher score meant better cognitive functioning.
ADL limitations score: ranging from 0 to 100 (full score). The measurement tool was Modified Barthel Index. Higher score meant better functional status.
IADL limitations score: ranging from 0 to 14 (full score). The measurement tool was the Instrumental Dependency-OARS. Higher score meant better instrumental functional status.
BPSD frequency score (equal to the total frequency score of 38 symptoms): Frequency score of each symptom ranged from 0 (no occurrence) to 6 (several times an hour). The measurement tool was the Adapted Cohen-Mansfield Agitation Inventory-Community Form. Higher score meant more frequent BPSD.
Carer stress score (equal to the total problem score): Each BPSD caused different levels of problems to carers, ranging from no problem (scored 0) to large problem (scored 4). Higher score meant higher care stress level.
We used PASW 19.0 to perform all analyses. Descriptive data were presented using mean, minimum and maximum figures, standard deviation and proportions. We performed Cox proportional hazards regression analyses (backward step-wise) for six, nine and 16 months’ study periods respectively to examine the short-term and long-term effects of baseline variables on time to death or hospital admission.
The Cox regression analyses included two steps. Step one was univariate analysis, aiming to identify potential significant factors for time to death or hospital admission of each study period (p-value set at 0.10). Step two was multivariate analysis involving running a Cox regression model for each study period by including all of its potential significant factors. This step determined the final significant factors for time to death or hospital admission for each study period (p-value set at 0.05).
Kaplan-Meier survival curve using censored data was displayed.
Results of the Cox regression analyses were shown in the form of hazard ratios. A hazard ratio of greater than 1 for a variable indicates that hazard increases as the value of the variable increases at any period of time, and vice versa for a hazard ratio of less than 1. The 95% confidence interval, and two-tailed P-value of 0.05 were adopted.
The sample size for this survival analysis was estimated based on 5% statistical significance, 80% power and minimum difference to be detected of 20% between categories in subgroup independent factors. It was unclear however at study outset what number of death and hospital admission events would occur in this population in the six, nine and 16-month study periods, as well as the size of the independent factor subgroups, all these parameters being necessary for sample size estimation.
Plausible estimates of numbers of these parameters indicated that sample size estimates were sensitive to variation of numbers used and whether the study would be adequately powered. This being so, death and hospital admission events were combined to increase the number of what became the principal outcome variable. Such combination is commonly performed in other fields of health care . Based upon the following plausible numbers — 15% death or hospital admissions, 70%/30% relative size of subgroups and 20% difference in subgroup effects, it was estimated that a total sample size of 237 was necessary. Assuming 20% attrition in numbers other than death and hospital admissions, the total sample size as required is 284.
Using the same method in sample size calculation but based on 10% death, 70%/30% relative size of subgroups and 10% difference, the total sample size of 433 (larger than the sample size of this study — 284) was necessary to examine risk factors for time to death. This further supports why we needed to combine death and hospital admission and primarily examined risk factors for time to death or hospital admission.