|
Abstract:
|
Controlled trials should compare like to like, so that differences in outcome may be confidently attributed to differences in treatment. Likeness, in a stochastic sense, is ensured by randomization and can be improved by adjustment. Likeness in any other sense is unnecessary, and attempts to improve it are usually useless and sometimes harmful. On the other hand, inference about subgroups may be important. The usual kind of stratification is sometimes thought to be essential to this purpose, but it is not. Other kinds of stratification, however, may be indispensable in this setting.
|
