Online Program

When Good Experiments Go Bad: A Case Study of Outliers (302832)

Keywords: outliers, linear models, sensitivity analysis, tumor volume

Data from a mouse study examining the effect of cisplatin or cisplatin+nutlin on bone tumor growth were analyzed conforming to standard lab procedures. While the data were completely separated between the two treatments and the control group, the initial analysis of variance (ANOVA) results were non-significant (P=0.091). Two control group tumor volumes were identified as outliers. To estimate tumor size within each group and treatment effect between groups, four methods were used: 1) ANOVA, 2) gamma generalized linear model (GLM), 3) Kruskal-Wallis (KW) nonparametric ANOVA and 4) robust regression (RW). Also, a sensitivity analysis compared all four methods excluding the outliers. ANOVA estimated negative tumor volumes for both treatment groups. GLM, KW, and RR each estimated strictly non-negative tumor volumes, and all three models indicated significant treatment effect (P<0.001). The sensitivity analysis confirmed the results of the GLM, KW, and RR methods. Compared to other methods, the ANOVA was observed to be sensitive to outliers. Exploring alternative estimation methods yielded correct inference without the exclusion of outliers.