Most comparisons of treatments or doses against a control are performed by the original Dunnett single step procedure \cite{Dunnett1955} providing both adjusted p-values and simultaneous confidence intervals for differences to the control. Motivated by power arguments, unbalanced designs with higher sample size in the control are recommended. When higher variance occur in the treatment of interest or in the control, the related per-pairs power is reduced, as expected. However, if the variance is increased in a non-affected treatment group, e.g. in the highest dose (which is highly significant), the per-pairs power is also reduced in the remaining treatment groups of interest. I.e., decisions about the significance of certain comparisons may be seriously distorted. To avoid this nasty property, three modifications for heterogeneous variances are compared by a simulation study with the original Dunnett procedure. For small and medium sample sizes, a Welch-type modification can be recommended. For medium to high sample sizes, the use of a sandwich estimator instead of the common mean square estimator is useful. Related CRAN packages are provided. Summarizing we recommend not to use the original Dunnett procedure in routine and replace it by a robust modification. Particular care is needed in small sample size studies.

翻译：大多数处理组或剂量组与对照组的比较均采用原始Dunnett单步法\cite{Dunnett1955}，该方法可提供校正p值及与对照组差异的同时置信区间。基于检验功效的考量，推荐采用对照组样本量更大的非平衡设计。当目标处理组或对照组出现较高方差时，相关配对比较的功效会如预期降低。然而，若方差在非受影响的处理组（如剂量最高且差异极显著的组）中增大，其余目标处理组的配对比较功效同样会降低。这意味着某些比较的显著性判断可能被严重扭曲。为避免这一不良特性，本研究通过模拟比较三种异质方差修正方法与原始Dunnett方法的性能。针对小样本和中样本量情况，推荐采用Welch型修正方法；针对中高样本量情况，使用三明治估计量替代常规均方估计量更具优势。相关CRAN软件包已提供。综上所述，我们建议在常规分析中摒弃原始Dunnett方法，改用稳健修正方法。在小样本研究中需特别谨慎。