OBJECTIVES: Traditional statistical choice of models in health care area may not be based on future predictive ability of the model but by statistical significance. Meanwhile, there is a concern on validity of ordinary least square (OLS) model because of heavily skewed or zero-filled data. In this study, we applied several methods to a sample data to evaluate the performances of different methods in the analysis of cost or length of stay (LOS) data. METHODS: The sample data was from a study to assess the health economic burden of patients with treatment resistant depression (TRD) in comparison to non-TRD patients among major depressive disorder (MDD) patients. Patient’s demographic and baseline characteristics were accessed. By using OLS, generalized linear models with different distributional assumptions (Gamma, Poisson, etc.) in one/two-part models, the differences of mean and confidence intervals of medical cost or LOS in TRD vs. non-TRD were estimated. Root mean squared errors (RMSE) or Pearson correlation coefficient (r) of models were calculated by five-fold cross validation. RESULTS: Among 28547 patients in the cohort, there were 3420 TRD vs. 25127 non-TRD patients. With Non-TRD group as reference, the differences of mean in medical cost ranged from around $3800 to $4500 by different models by bootstrapping. The difference of mean for LOS was around 0.4 days. OLS model showed the most comparable results between least square means and bootstrapping methods. The cross validation results showed RMSE or r were similar across models. In sensitivity analyses, RMSE decreased with increasing cutoff percentile as extremes in cost or LOS data. CONCLUSIONS: In current study, because of limitation of study model or population, the cross validated results didn’t show much difference across models. This study will initiate further investigations regarding comprehensive predictive factors selection or using more powerful statistical methods with better fit to the data.