Mineral processing engineers often conduct trials to improve the performance of their plants. A common problem in such trials is detecting real but relatively small improvements or changes in process performance against a background of very noisy data. This large data variance is frequently caused by (among other things) long-term time trends in performance, which can be a result of systematic changes in feed condition. This paper describes two well-known statistical procedures for dealing with such situations, the paired t-test and the randomised block experiment. These methods are illustrated through their application to three real case studies in base metal flotation plants, all involving ＇＇yes-no＇＇ decisions, and all using metal recovery as the main performance criterion: 1. The evaluation of a new flotation collector in a production plant. 2. The assessment of two alternative flotation circuit configurations in a pilot plant. 3. Determination of the value of introducing a regrind stage ahead of a flotation circuit in a production plant. The paper considers the practical problems encountered in these experiments, discusses the compromises sometimes required in analysing imperfect experiments, and shows how the statistical procedures can be used to make good decisions in the face of uncertainty. The formulae and computational procedures are given in full to encourage their application to similar situation in the practice of mineral processing.