There is a constant pressure to provide statistical designs that make best use of limited money, time, and patient recruitment resources. Balancing the needs of the researcher with the well being of the patients is a constant challenge. Ten years ago when I started my job I had a lot of people coming to me with their data and asking (in so many words) “What do I do now?” Though many of these were student and resident researchers, this is no longer the case. IRB’s have become much stricter about requiring sound statistical plans even for student research, and journals are much tougher in their statistical reviews. The result is better use of resources, better use of the available patients for study, and better science leading to improved patient care.

@ J. Venier: You are absolutely correct that the most efficient design for a linear relationship is to sample at the two farthest extremes, but this design cannot detect any non-linear relationships. If you don’t know what so relationship to expect, the you may need to sample in between as well.

I think your point about statistical power is very important and deserves amplification. I’m no expert in experimental design, but I think it is pretty clear that efficient designs from a statistical standpoint often involve extreme conditions. For example, if you want to experimentally determine the relationship between altitude and air pressure, but could only make a limited number of observations, you will almost certainly want to sample near the ground and at the highest spot you could reach, to start with. The best elevations for additional observations depend on your assumptions, but if you assume a linear relationship and just want to determine the slope, I think it would be best to put half of your observations on the ground and half at the highest spot available.

But for example when testing how much of a drug people can tolerate, one of the analagous extreme conditions would be giving someone a dose which he is very unlikely to tolerate. Clearly, if you could do an experiment which is almost as good but only involves doses which are much more likely to be tolerated, this would be a more ethical choice, even if it is not as good from a statistical standpoint.

Anyway, you sure said a mouthful when you said that ethical constraints make biostatistics interesting! Especially since the ethical considerations often involve extremely deep questions.