Ultrasound brain scanning in the newborn.

patients, and in these the risk of imbalance is much greater. How should imbalance be dealt with? The crucial question is whether the variable in question is related to the outcome being studied. If, for example, one treatment group is on average five years older than the other this will be of varying importance for outcomes such as survival time, reduction in blood pressure, or response to treatment for eczema. In the first instance both clinical knowledge and simple statistical analysis may be used to identify potentially important imbalances, and various statistical approaches can be used to investigate further. A more advanced statistical analysis, using modelling techniques, may be needed, in which case, as Lavori et al suggest, expert statistical help is recommended.2 Only one of the 47 trials they reviewed had adopted this approach; they thought that over half could have benefited in this way. If there is a strongly prognostic variable there are good grounds for using the modelling approach even when perfect balance is achieved.26 In many cases some variables of prognostic importance can be identified when a clinical trial is being planned. Various techniques exist for ensuring balance between groups for important variables, the best known being stratified randomisation.2 7 Such methods may be especially beneficial in small trials or where a particular variable is known to have a powerful influence on outcome. DOUGLAS G ALTMAN Statistician, Clinical Research Centre, Harrow, Middlesex HA1 3UJ