Randomization or the conducting of Experimental Runs in random order provides the foundation for making valid comparisons among the levels of the experimental factor. The basic assumption underlying such comparisons is that all other Factors and conditions in the experiment are constant, except for the experimental factor under study. If an environmental condition changes during an experiment, and the runs are not randomized, the results may be biased by the environmental (lurking variable) shift. Thus, randomization helps ‘even out’ the impact of Lurking Variables, insuring the fairness of the comparisons.
Consider an experiment to evaluate the effect of a new treatment for migraines. In such a study randomization can occur in two ways: the first is the random selection of the subjects or units to be included in the study. Here random selection is important to define the population to which the study results can later be generalized.
The second randomization is the random assignment of the subjects to the treatment or control groups respectively. This strategy helps insure against bias due to one type of subject being assigned to the treatment group more often than another (such as patients suffering from worse migraines being subconsciously assigned more often to the new treatment). If, in general, younger subjects tend to suffer worse migraines than older subjects, then a non-randomized study would confound the true treatment effect with the age effect, where age is a lurking variable.
Randomization in Designed Experiments by Keith M. Bower. ASQ Six Sigma Forum, December 2004. - http://www.box.net/shared/static/a5g3b3gs8o.pdf