OBJECTIVES: Economic modeling and healthcare cost analyses are used to inform policy-makers in health care decision-making such as cost-of-illness assessments, treatment evaluation studies, and more commonly to predict healthcare costs for specific patient populations. However, due to stochastic error distribution assumptions and the challenges they create for econometric modeling, various types of models have been identified to address specific distributional characteristics.  The objective of study is to examine cost distributions for treated diabetes patients and compare untransformed ordinary least squares (OLS) regression with log transformed OLS and generalized linear model (GLM) methods.  METHODS: A simulated distribution was created mirroring a representative claims dataset for type 2 diabetes costs.  Using simulated cost distributions with known error term ensures that model selection could be assessed with certainty of the error specification for the distribution.  Two simulated cost distributions were generated: one with homoskedastic errors and the other with heteroskedastic errors.  Both distributions were used to explore model performance under varying conditions.  Several tests for model fit, specification, and predictive ability were selected from the existing literature to assess model selection and determine best model fit.  Simulation and all analyses were done using STATA 11. RESULTS: Results from the model specification tests indicate that for both cost error distributions, OLS regression on untransformed Y is the best-fitting model of the three tested.   Although superior in model fit, prediction criteria indicate that OLS is relatively poor in prediction along the entire range of costs.    CONCLUSIONS: OLS on untransformed Y cost model is selected as the best model choice under the conditions of the given distributions.  The inability for all three models to predict within the full range of costs demonstrates weakness in characterizing the upper tail of the distribution.  If prediction is the primary concern, a two-part model may be more appropriate.