The problems that arise in the statistical analysis of feeding trials are more often due to mistaken philosophy than to particular technical difficulties related to the analysis of the trials. For example, the title originally proposed for this paper was 'Problems of applying statistical designs and analysis to feeding trials', which suggests that the nutritionist would happily conduct his trials in isolation, were it not for the need to adapt them to satisfy the peculiar demands of the statistician. The problem that in fact we face is to design feeding trials in such a way that they will produce truly meaningful results which are of use to the rest of the world.At its simplest, an experiment in applied science asks whether treatment A differs from treatment B in its effect on some measurable character. The experimenter then tests A and B on representative samples of the population to which the result is to be applied. By statistical analysis he can compare the difference obtained with the variation between items making up the samples and, after making some assumptions about the distribution of this variation, may be able to claim that if the effects of A and B are truly the same he has witnessed an unlikely event, and he can say how unlikely. It is then up to the experimenter and his readers to decide whether they believe that an unlikely event has occurred or that there is a true difference.The theoretical concepts of truth and representative sample lead to great difficulty in interpreting and hence designing feeding trials. The nearest that statisticians can get to defining truth is to call the true (or expected) value the average over the whole population, and for them the ideal material is an infinite population of random variates, which will obviously differ from a real population of animals with which we have to deal. Similarly, a truly representative sample is a random sample which is drawn rigorously so that every item has the same chance of being selected, but it is much simpler conceptually and practically to generate a set of random numbers from a particular distribution using a computer than it is to select a random sample of pigs intended to represent all pigs now present or likely to be bred in the foreseeable future. The nature of the sample of animals probably accounts for many of the discrepancies between results from different centres. The best that can be said of feeding trials is that the results refer to the animals used and any other population of which they could be considered a random sample, but the latter is difficult to assess.This problem becomes particularly acute where, for purposes of economy, experiments are carried out on an institute's own herd of animals. This practice at https://www.cambridge.org/core/terms. https://doi