I'm not involved in drug testing, but it is my understanding that most trials are performed using a control and a treatment group (or perhaps a few discrete treatment groups). This data is then used to determine effectiveness of a drug, and hopefully uncover any deleterious side-effects before the drug is released on the population.
I'm wondering why it would not be more sensible to use a continuous distribution of drug dosages, rather than a discrete distribution? For example, give 20% of patients the control, and the give the other 80% a randomised (uniform?) dosage between 0% and ~110% of the expected optimal dosage. It would seem to me that this would provide more information than data that is inherently discretised, as well as potentially making it easier to more accurately fit models like logistic functions for both effects and multiple side-effects that occur at different rates?
The obvious draw-back would be the increased cost of manufacture for the trial, but that would be fairly minimal in comparison to potential benefits, especially for large trials. Also, this method could just as easily be applied to any "treatment" experiment (e.g. in agriculture, metalurgy, etc.). So why aren't more trials conducted like this?