OBJECTIVES: There is controversy regarding the most appropriate method to analyze cost data. We examined the role of censoring, model fit and assumption violations in a study determining the costs of dialysis modality for the treatment of kidney failure. METHODS: All incident dialysis patients from Alberta from 1999 to 2003 were tracked for direct medical costs (inpatient and outpatient costs, physician claims, and medication costs) for three years using administrative datasets, and categorized into peritoneal dialysis (PD) only, hemodialysis (HD) only, or modality switches (HD to PD; PD to HD). Unadjusted censored and uncensored cumulative costs were determined using the non-parametric method of Lin and bootstrapping.  Model fit and assumption violations for alternate covariate adjusted regression models were assessed through comparison of various approaches (OLS or GLM regression, log-transformation, smearing), and the optimal approach identified. RESULTS: Three year adjusted cumulative costs for patients receiving PD only and HD to PD groups were $58,724 ($44,123-$73,325; 95% confidence interval) and $114,503 ($96,318-$132,688) respectively, and were significantly lower than patients receiving HD only $175,996 ($134,787-$217,205) and PD technique failure patients $173,308 ($147,725-$119,891).  Comparison of censored and uncensored unadjusted total cumulative costs yielded similar results.  Covariate adjusted GLM regression was the best fit for modeling total cumulative costs when compared to other OLS and GLM models.  For cost categories with smaller sample sizes (inpatient costs; n=674), GLM models fitted poorly due to kurtosis, and OLS regression on log-transformed costs with smearing yielded better less biased estimates.  CONCLUSIONS: GLM regression methods for estimating costs performed well when applied to our analyses with larger datasets (n>1000), but gave biased results when smaller datasets were used.  When modeling costs with smaller sample sizes, the most appropriate regression method can be determined through model diagnostics.