The three-arm design with test treatment, reference treatment and placebo offers internal assay sensitivity for a proof of non-inferiority. In this design the relative effects known from nonparametric theory are robust tools allowing the assessment of non-inferiority in a range of situations. An asymptotic nonparametric theory is established in the three-arm design based on the asymptotic distribution of rank means under alternative. A rank test for non-inferiority is derived. Fieller's formula is used to calculate a respective confidence interval. The approach is expanded to multicentre studies. The simulation studies are conducted demonstrating the accuracy of the methods and an example is discussed.