Operating metallurgists spend a significant part of their time running plant trials to compare two or more conditions in order to improve the process, such as a new reagent, flow sheet or piece of equipment. The best way to conduct such a trial is as a paired experiment when testing two conditions (eg 'old' and 'new') or a randomised block design when testing more than two conditions. These experimental designs should always be preferred over alternatives if their requirements can be met, because they block the effect of uncontrolled variables ('covariates') and are the most statistically efficient, ie they will lead to the required level of confidence in the outcome faster than other methods._x000D_
Sometimes, however, it is not possible to switch condition easily, which is a prime requirement of these methods. In such cases the only option may be to switch condition once only, and compare the process performance before and after the switch. Examples are testing grinding media, large equipment, or long residence time processes such as leach trains or CCD thickeners. This paper considers some data analysis methods to deal with this situation. Modelling with intervention analysis (IA) uses regression models with a dummy variable to indicate the presence or absence of the test condition. IA with a time series model ensures that the residuals are uncorrelated. IA with a process model helps to deal with the problem of covariates. The reference distribution is a method free of any statistical assumptions but is less powerful than the other methods._x000D_
Cusum charts are a helpful visual aid to interpreting time trends. A combination of these methods can help improve the confidence in the final decision. None are as good as a formal experimental design such as a paired trial._x000D_
FORMAL CITATION:Napier-Munn, T J, 2008. Practical problems in running statistically valid plant trials, and their solution, in Proceedings MetPlant 2008, pp 249-258 (The Australasian Institute of Mining and Metallurgy: Melbourne).