A Tutorial on What to Do With Skewness, Kurtosis, and Outliers: New Insights to Help Scholars Conduct and Defend Their Research
通过模拟研究，考察偏度、峰度和异常值对常用统计模型的影响，发现单变量非正态数据影响不大，但多变量异常值需谨慎处理。
ABSTRACT This research examines the effect of nonnormal data on several frequently utilized statistical models. Behavioral scholars often encounter nonnormal and potentially problematic data such as outliers. In the absence of clear guidance, scholars generally report taking some compensatory action, including the deletion of data, data transformations, or alternative statistics. This research addresses the question, how much do such data actually affect results? Three sets of simulations investigate the effects of skewness, kurtosis, and outliers. Our findings show that some patterns of nonnormal data are perhaps not as problematic as one might anticipate. Many results are affected but in the statistically conservative direction (i.e., less likely to produce significance). In particular, data that were skewed, or showed kurtosis or univariate outliers did not lead to inappropriate statistical results, and, therefore, scholars may proceed with their analyses. Yet there are boundary conditions; specifically, we found that multivariate outliers could create biased results, so for such data, scholars need to be particularly transparent and vigilant in their data handling, posting data sets with and without any deleted or transformed data points.