OBJECTIVES: To illustrate how use of generalized linear models to analyze health care cost data may provide a better distributional fit than commonly employed approaches (e.g., linear or log-linear ordinary least squares [OLS] regression), and could yield quantitatively and inferentially different conclusions. METHODS: Data were obtained from the PharMetrics Patient-Centric Database, which includes integrated medical pharmacy claims from 73 health plans nationwide. Patients with a diagnosis of intermittent claudication (ICD-9-CM 443.9x) who newly started cilostazol or pentoxifylline therapy between June 1999 - March 2002 were selected for analysis. Six-month pretreatment and follow-up periods were created in relation to the first observed prescription. Total costs of care during follow-up were estimated based on health plan payments for medications and services rendered, and were expressed in 2002 U.S. dollars. Alternative multivariate approaches to analyzing total costs were employed-an OLS model (log-linear) versus a generalized linear model (GLM) with a log-link function and a gamma distribution. Covariates included demographic and other baseline/pretreatment variables. Histograms of untransformed and log-transformed costs were compared to gamma and normal distributions; goodness-of-fit assessments also were conducted. The results of OLS (on a log-transformed outcome) and gamma GLM models were compared. RESULTS: Analyses were conducted for 763 and 506 patients newly starting cilostazol and pentoxifylline therapy respectively. The results of goodness-of-fit testing (deviance: 1489.5 vs. 1366.4 for degrees of freedom = 1,255) indicated that the gamma GLM model approximated the cost distribution most closely. Observed annual mean total costs were $6238 and $5568 for cilostazol and pentoxifylline respectively; application of the two models yielded different results-a nonsignficant (p = 0.0620) treatment effect using log-linear OLS, and significantly (p = 0.0432) higher costs for cilostazol using a gamma GLM model. CONCLUSIONS: The gamma GLM technique is a powerful tool for modeling strictly positive skewed outcomes and should be more widely employed in pharmacoeconomic analyses.