Keywords: Cost data, Relative risk, Semicontinuous data
Cost data often contain many zero values. Two-part models with separate models for part I, probability of non-zero cost (i.e. use of service) and for part II, amount of non-zero cost (i.e. cost of service) are often used. Total marginal per patient cost is due to a combination of probability of use and mean cost of use. Previous two-part models quantify the relative contributions of these two sources to total cost conditional on a specified covariate pattern, complicating interpretation. This paper presents two approaches that provide simple decompositions of covariate effects on total cost into these two components on comparable scales and thus quantify their relative contributions to total cost. Simulations illustrate interpretation and validity of the methods across a range of conditions. The methods are applied to risk adjusted 30-day outpatient cardiac cost for patients following percutaneous coronary intervention (PCI) from the Department of Veterans Affairs (VA) Clinical Assessment, Reporting and Tracking (CART) program.