Introduction to Chapter 2: The Design of Experiments

Extract from Statistical Questions in Evidence-based Medicine by Martin Bland and Janet Peacock.

Each chapter of Statistical Question in Evidence-based Medicine starts with a one-page introduction, which describes the statistical methods which will be covered in the chapter. We hope that the topic will be useful in own right, as well as giving a flavour of the book.

This chapter covers the design of experiments, in particular clinical trials on human patients.

In an experiment or trial we carry out a treatment and observe the outcome. We need a control group who receive no treatment or a different treatment, but are otherwise similar to the group receiving the treatment of interest. The groups of patients should be comparable in every respect except treatment. We achieve this by random allocation, where subjects are allocated to treatments without the subject's own characteristics influencing the choice. Each subject has the same chance of receiving each treatment. Methods to do this include tossing a coin and random number tables.

Experimenters' observations may be influenced by knowledge of which treatment a subject has received. If possible, we should conceal treatment from the observer, a procedure called blinding or masking. Subject response also may be influenced by knowing which treatment they have received, so trial subjects, too, should be blinded if possible. `Double-blind' means that the neither the subject nor the experimenter knows the treatment allocation. Where there is no control treatment, we may use a `placebo', which is an inactive treatment, outwardly as similar as possible to the active treatment.

A cross-over trial is one where each subject receives all treatments, in random order. Each subject acts as their own control. This is good for treatments which control symptoms in chronic disease. We need a smaller sample size than for a two-sample trial, but we cannot look at long-term effects of treatments or at treatments which cure the disease.

Treatment groups are comparable at randomization, but this may change if some patients drop out or switch treatments. To keep the groups comparable, we analyse according to the original intention to treat, including everybody as if they had received the treatment to which they were allocated, whether they actually did or not.

Sometimes trial subjects are allocated into treatments in clusters, e.g. all patients in a general practice. The cluster is the experimental unit. Data must be analysed in experimental units, not as individual subjects.

It is essential that trials on human subjects be carried out ethically, the welfare of the trial subject being the researchers' primary concern. Potential subjects should be informed as fully as possible of what is involved in the trial and their free consent obtained.

Back to Statistical Questions in Evidence-based Medicine contents.

Back to Martin Bland's home page.

This page maintained by Martin Bland.
Last updated: 30 June, 2005.

Back to top.