Seminars

The classical random and fixed effects models rely on the assumptions of normality and homoscedasticity. In some designs, these assumptions (and some additional ones) are needed to guarantee the assumed independence of the random effects. The fully nonparametric formulation, which will be introduced, does not require these assumptions. In particular, it pertains to both discrete and continuous ordinal data, and the distribution of the response variable can depend in an arbitrary fashion on the levels of the random and fixed effects. Test procedures that rely on high dimensional low sample size asymptotics are developed and compared with the classical F and Hotelling's T-square statistics. The new procedures are illustrated on a data set from the NOAA mussel watch project.