Background: Analysis of the patterns of variation in health care costs and the determinants of these costs (including treatment differences) is an increasingly important aspect of research into the performance of mental health services.
Aims: To encourage both investigators of the variation in health care costs and the consumers of their investigations to think more critically about the precise aims of these investigations and the choice of statistical methods appropriate to achieve them.
Method: We briefly describe examples of regression models that might be of use in the prediction of mental health costs and how one might choose which one to use for a particular research project.
Conclusions: If the investigators are primarily interested in explanatory mechanisms then they should seriously consider generalised linear models (but with careful attention being paid to the appropriate error distribution). Further insight is likely to be gained through the use of two-part models. For prediction we recommend regression on raw costs using ordinary least-square methods. Whatever method is used, investigators should consider how robust their methods might be to incorrect distributional assumptions (particularly in small samples) and they should not automatically assume that methods such as bootstrapping will allow them to ignore these problems.