Page 87 - Reduction of coercive measures

measures aimed to prevent from indirect danger or disadvantage. The third subset contained 6 coercive measures resulting from the use of surveillance technology. The fourth subset contained 10 coercive measures resulting from the of use of material to physically support the resident. Hypotheses concerned the sum of all applied coercive measures, the sum of coercive measures applied at direct danger, and the sum of coercive measures preventing from indirect danger or disadvantage. Per resident it was calculated how many measures of each type of measure were registered at that moment. Appendix A contains the full list of coercive measures by subset and their frequency.
Statistical analysis
Associations between characteristics of residents and support staff and the use of coercive measures were tested by using generalized linear mixed modelling in SPSS version 24. This modelling technique accounts for the dependency of observations due to the multilevel structure of the data (residents nested within units) (Hox, Moerbeek, & van der Schoot, 2017). Dependency among the factors related to unit staff was addressed by averaging scores from staff belonging to the unit of each resident. Because the dependent variable was a count variable, a negative binomial regression analysis was conducted, which uses a log function to link the dependent count variable to the independent variables in the model. This model was deemed more adequate than the Poisson regression model that can also be applied to count data, because the variance of the count variable ‘total number of coercive measures’ was larger than its mean (overdispersion). The analysis were conducted in several steps. First, a so-called ‘empty’ model was run in which the nested data structure was specified but no predictors were included. This model allows to estimate the intraclass correlation (ICC), which is the proportion of variance in the count variable that can be attributed to the level of the unit. This correlation indicates to what extent residents within the same units resemble one another on ‘total number of coercive measures’. A rule of thumb is that if the ICC is larger than .05 the multilevel data structure cannot be ignored and mixed modelling is indicated (Hox
Associate factors of coercive measures
85
4