THE NEED FOR RANDOM SELECTION

Frequently in this book we have talked about ‘random’ samples and ‘random’ selection. The word ‘random’ has been used in a more or less intuitive way to imply the purely haphazard and unbiased collection of measurements or allocation of treatments to the experimental material. At the same time, it has been hinted that it is often necessary to ensure that the events in question are sufficiently random by appealing to some special technique. Although one can easily get into deep philosophical waters over questions of the ‘true meaning of randomness’, it is unnecessary to broach such difficulties here. So far as the experimenter is concerned, he or she wants to avoid the sort of bias that will lead him or her to infer, for instance, real differences between treatments, when the observed differences are due merely to some heterogeneity in the test material. We shall therefore discuss randomness only in relation to avoiding various forms of conscious and unconscious bias. The use of tables of random numbers is the chief means of achieving this object. (It is, however, worth mentioning in passing that randomisation in properly designed experiments, such as those with randomised blocks, for example, does in fact make the usual significance tests approximately valid even when the assumption of normally distributed residual variation does not hold.)

The possibility of simple forms of bias will of course occur quite naturally to the competent experimenter.