Abstract:
|
The hurdle model is a well-established method for modelling zero heavy data. Common examples include counts of dental caries, hospital length of stay, and monetary expenditures. In a study of one-year post discharge medical costs (beyond routine clinic visits) it was observed that patients had, perhaps, more than one threshold. Some had a relatively low cost which may have included an emergency room visit, but there was a distinctly costlier group which was composed of those readmitted to the hospital. To fit these data, we propose an adaptation of the hurdle model that uses multinomial regression to model covariate effects on the outcome of cost group, and separate gamma regressions to model the mean cost once a cost is realized. Specific challenges as well solutions will be discussed including using multiple parameter vectors in the model, non-linearity of parameters, and interpretation by estimation of contrasts at mean covariate values and comparisons of mean predicted values.
|