1 Forskningsenheden for Almen Praksis, Faculty of Health Sciences, Aarhus University, Aarhus University2 Department of Public Health - Forskningsenheden for Almen Praksis, Department of Public Health, Health, Aarhus University3 Forskningsenheden for Funktionelle Lidelser4 unknown5 Department of Public Health - Forskningsenheden for Almen Praksis, Department of Public Health, Health, Aarhus University
We consider health care data from a cluster-randomized intervention study in primary care to test whether the average health care costs among study patients differ between the two groups. The problems of analysing cost data are that most data are severely skewed. Median instead of mean is commonly used for skewed distributions. For health care data, however, we need to recover the total cost in a given patient population. Thus, we focus, on making inferences on population means. Furthermore, a problem of clustered data is added as data related to patients in primary care are organized in clusters of general practices. There have been suggestions to apply different methods, e.g., the non-parametric bootstrap, to highly skewed data from pragmatic randomized trials without clusters, but there is very little information about how to analyse skewed data from cluster-randomized trials. Many studies have used non-valid analysis of skewed data. We propose two different methods to compare mean cost in two groups. Firstly, we use a non-parametric bootstrap method where the re-sampling takes place on two levels in order to take into account the cluster effect. Secondly, we proceed with a log-transformation of the cost data and apply the normal theory on these data. Again we try to account for the cluster effect. The performance of these two methods is investigated in a simulation study. The advantages and disadvantages of the different approaches are discussed.