Technical Report / Universität Dortmund, SFB 475 Komplexitätsreduktion in Multivariaten Datenstrukturen 2004,47
In sensory experiments, often designs are used that are balanced for carryover effects. It is hoped that this controls for possible carryover effects, like, e.g., a lingering taste of the products. Proper randomization is essential to guarantee the usual model assumption of independent identically distributed (i.i.d.) errors. We consider a randomization procedure that permutes treatment labels and assessors. This restricted randomization leaves the neighbour structure unchanged and validates the assumption of i.i.d. errors if the design used is a Generalized Youden Design (GYD). However, the use of a neighbour balanced GYD may require too many assessors. The question arises, whether nearly balanced designs may be used without grossly violating the validity of the analysis. We therefore do a simulation study to assess the properties (under this restricted randomization) of nearly balanced designs like, e.g., the ones proposed by Périnel and Pag?s (2004, Food Quality and Preference 15, 439?446). We observe that, if there are no carryover effects, the variance estimates for treatment contrasts are not significantly biased whenever we use designs that are nearly GYD. Additionally, designs that are nearly carryover balanced still produce conservative variance estimates, even in the presence of large carryover effects. In all, ?nearly neighbour balanced nearly GYD? as proposed by Périnel and Pag?s (2004) appear to be useful in experimental situations where the use of GYD is too restrictive. It should be stressed, however, that these results are true only if randomization is used as a protection against effects unaccounted for in the statistical model.