A scalar model output Y is assumed to depend deterministically on a set of stochastically independent input vectors of different dimensions. The composition of the variance of Y is considered; variance components of particular relevance for uncertainty analysis are identified. Several analysis of variance designs for estimation of these variance components are discussed. Classical normal-model theory can suggest optimal designs. The designs can be implemented with various sampling methods: ordinary random sampling, latin hypercube sampling and scrambled quasi-random sampling. Some combinations of design and sampling method are compared in two small-scale numerical experiments.