The difference in arithmetic mean cost is the primary cost outcome of interest in cost-effectiveness analysis. Its analysis is complicated by the routine nonnormality of cost data which violates assumptions of most parametric techniques used to analyze raw cost data. Commonly proposed methods to address this issue--nonparametric tests like the Wilcoxon rank sum test or transformations of the data--are inappropriate because they test something other than the arithmetic mean or pose difficulties for estimation of the predicted difference in arithmetic mean cost. Use of a nonparametric bootstrap (univariate) or estimation techniques such as generalized linear models (multivariable) allows estimation of and the drawing of inferences for the difference in arithmetic mean. Problems with the more commonly proposed solutions and the alternative methods that avoid these problems are described.