OBJECTIVES: It is generally agreed that calculation of means after non-linear data transformations (e.g., log-transformation) does not result in a comparison of arithmetic means, and so is not appropriate for cost data in pharmacoeconomic evaluations. This would seem to preclude the use of log-normal distributions for cost data in probabilistic sensitivity analysis. The study objective was to investigate the statistical properties of arithmetic mean costs. METHODS: Monte Carlo simulations were use to investigate the statistical properties of arithmetic mean costs derived from an underlying log-normal distribution, loge (X) ~ N(m,s2), where m = loge(€10), s =1.5 (range 0.5 to 2.5). An underlying log-normal distribution was used because cost data are typically highly positively skewed. Microsoft Excel was used to perform the Monte Carlo simulations generating 1,000 arithmetic means, each from a sample of N=100, for each value of s investigated. RESULTS: The distribution of arithmetic means increased in positive skewness as s increased. For s ≥ 1.5, the distribution of arithmetic means deviated considerably from normality. The level of skewness was greatly reduced by use of the log-normal distribution. The Gamma distribution was similar to the log-normal distribution in representing the distribution of arithmetic mean costs. CONCLUSIONS: Log-normal distributions for arithmetic mean cost data may have a role for use in probabilistic sensitivity analysis, although this needs further investigation using cost data derived from actual studies.