The present study is a retrospective observational study that aims to evaluate health effects of reablement service in a municipality.
2.1 Description of the data
The context of this study is a relatively small, suburban municipality in Norway with less than 15 000 inhabitants. Reablement was implemented into home-care services seven years prior to this study.
Data isolated to the reablement service was not available from the national registries on health care services in Norway. To collect data, we had to extract data directly from the patient administrative system of the municipality included in this study. The system contained all users that had applied for, or received, municipal health or care services in that municipality in the period 1993 to 2018. From this we extracted the history of use of services and individual needs for care. Need for care is measured for each user that applies for municipal health and care services and is measured on activities of daily living functional scales.
The data collected contained 4 872 individuals, however this included also test data, deceased individuals, and people who had moved from the municipality. Thus, the real number of health and care users in the municipality was less than 1000.
To select the reablement users and the control group some users where excluded. These were users without any services registered, users with no relevant services registered (i.e. parking permissions), users with services stopped prior to 1.1.2014, users without registered need for care (ADL measurement), users without need for care (ADL measurement) within the time frame of the study, and finally we also excluded users whose first service in the municipality was not an at-home service. See figure 1 for inclusion and exclusion flow chart. This data washing resulted in 153 reablement users and 697 without reablement.
2.2 Methods
2.2.1 Survival analysis
The purpose of survival analysis is to analyse the possibility of an unwanted event by comparing different groups under risk. The main outcome variable in survival analysis is time until an event occurs (26), typically referred to as survival time. The unwanted event can be any number of different outcomes that may happen to an individual.
Our study examines two outcomes: mortality and long-term care. Both outcomes are operationalized in two ways depending on the type of analysis.
Date of death was registered in the patient administrative system for the diseased. For the survival analyses we measured time to death as the number of days from the first registered service to death. In the regressions comparing the reablement and non-reablement groups, the outcome is operationalized as a dummy, i.e., death or survival.
Long-term institutional care is a level of service for users in the municipality. This will typically include nursing homes, but in some municipalities, this is offered instead as sheltered housing. Sheltered housing tends to be less staffed than nursing homes and thus come with lower average costs. The outcome in our analyses was operationalised twofold: 1) as a dummy (yes/no) if the user receives long term care at any point, and 2) the number of days from the first registered service to the possible unwanted outcome.
In the analysis we will use two main methods. The first is to produce Kaplan-Meier curves. These curves display the cumulative probability of survival at any given point in time after the start of the study. Survival is not limited only to avoiding death, but also avoiding any other predefined unwanted event. In the present study we do look at death and examine the use of long-term care as unwanted outcomes, and survival is interpreted as avoiding these outcomes over time. The Kaplan-Meier curves will indicate if the unwanted event occurs sooner or later for the different groups. The best possible outcome as time passes (i.e. towards the right-hand side of each graph) is for each group to have a large share surviving – i.e. not showing a declining curve. The Kaplan-Meier graph will then show two important traits; 1) what the end result is, that is; which group has the highest level at the end of the curve i.e. the longest time before an unwanted event occurs, and 2) how the development over time is for each group.
We will also estimate Cox proportional hazard model. This is a mathematical method for analysing survival data. The method includes multivariate regressions to describe hazard ratios of the different variables, estimated using Maximum likelihood. Hazard is the potential for the unwanted event to occur, given the survival time. Cox proportional hazard models are reliable and robust for analysing survival data (26).
2.2.2 Propensity score matching
In this paper we test the outcome of one group of patients: reablement users. This is one specific service in one specific municipality. If we compare just the average of this service to the average of those that do not receive this service, we are likely to find large differences due to selection biases. In our setting, it was not possible to achieve randomization. To overcome this, we employed a matching strategy.
For each user of reablement, we matched a home care user who did not receive reablement services. Propensity score matching (27) was the method applied to obtain the best possible match. By comparing similar age, sex, and level of ADL functioning, the method constructs as close a match as possible, so that the starting point for comparison is not all that different. Technically, propensity score matching involves a logistic regression to calculate the probability of a unit belonging to the treatment group (reablement) based on the included control variables (age, sex, and functionality level). The results after the matching were a set of weights that balanced the dataset so that the treatment and non-treatment groups can be matched and compared.
In the matching process we also estimated the average treatment effect which is the average difference between the reablement and regular home care users conditioning on the covariates (age, sex, and functionality level).
After matching we performed a simple OLS regression to test if there were differences between the reablement and non-reablement groups, even after both matching and controlling for age, sex, and functionality level.