Multilevel modelling of cost data: an application to thrombolysis and primary angioplasty in the UK NHS

Cost data are frequently collected from several locations and tend to be non negative and skewed. Generalised linear multilevel models provide a means of dealing with each of these issues. This paper compares several statistical models within this class using data drawn from an observational study of 3,000 patients treated for heart attack in 15 UK NHS hospitals. A number of alternative link functions and covariates were considered. We demonstrate that whilst it is important to take account of clustering in the data, the precise manner in which this is done is equally important. Models which allow for correlation between the random effects components and heteroskedasticity across all hospitals performed best in terms of model fit and made substantial di¤erences to cost estimates.