Summary: In a fundamental critique, E. v. Collani [Metrika 38, No. 1, 19-36 (1991; Zbl 0714.62094); H.-J. Lenz et al. (eds.), Frontiers in Statistical Quality Control, 91-99 (1992; Zbl 0789.62081] argued that acceptance sampling by variables is inappropriate if one is interested in the fraction nonconforming in incoming batches. We analyze this critique from a decision theoretic point of view, using a super-population model and loss structures that depend on the realization of individual lots. For important cost models, sampling by variables is always optimal. For other cost models it is compared to sampling by attributes; it is much better for large lots, but may be inferior for small ones. A mixed strategy is proposed; for a suitable choice of parameters, it behaves like variables sampling for large batches and automatically includes more elements of attributes sampling for small ones.