The Type I error of the Kenward-Roger (KR) F-test was assessed through a simulation study for a between- by within-subjects split-plot design with non-normal ignorable missing data. The KR-test for the between- and within-subjects main effect was robust under all simulation variables investigated and when the data were missing completely at random (MCAR). This continued to hold for the between-subjects main effect when data were missing at random (MAR). For the interaction, the KR F-test performed fairly well at controlling Type I under MCAR and the simulation variables investigated. However, under MAR, the KR F-test for the interaction only provided acceptable Type I error when the within-subjects factor was set at 3 and 5% missing data.
Padilla, Miguel A.; Min, YoungKyoung; and Zhang, Guili
"Type I Error Rates of the Kenward-Roger F-test for a Split-Plot Design with Missing Values and Non-Normal Data,"
Journal of Modern Applied Statistical Methods:
2, Article 4.
Available at: http://digitalcommons.wayne.edu/jmasm/vol7/iss2/4