Statistical principles over-design

As with all research involving humans and other intelligent creatures, it is also essential to design the experiments to avoid the introduction of unintentional bias from over-helpful subjects. We should remember the example of geneticist Mendel and his overly helpful fellow monks, who quietly weeded the pea patch so that the results came out in line with Mendel s predictions. Thus, where appropriate, variants on the double-bfind principle should be apphed in experimental design, and statistically significant numbers of measurements need to be made to properly account for the inherent variability of living subjects. 

The simplex method has been used widely over the past 30 years, its success as much owing to its simplicity as to its efficiency. Unlike the other methods described in this chapter and most of the others in this book, it assumes no mathematical model for the phenomenon or phenomena being studied. The often long and costly phase of determination of a model equation may therefore be avoided, and the method is thus economical in principle. It is sequential because the experiments are analysed one by one, as each is carried out. Because the method is not model-based we will not describe it in detail, but well indicate how it can "fit in" and complement statistical experimental design. 

In principle the method is simple enough - small quantities of the two calibrants in turn are allowed to equilibrate 30 °C below the temperature of the transition as indicated by a constant instrument signal. The calibrants are then heated through the transition and the extrapolated onset temperatures obtained from the endotherms. These temperatures correspond to Texpi and rexp2- The calibration will be affected by heating rate and possibly the nature of the atmosphere and its flow rate. The experimental conditions for the calibration should be chosen to match those for subsequent measurements. Calibration over more than two points may be carried out and the relationship between T and T xp determined statistically. The extent to which the instrument output can be corrected by the software will depend on the detailed design of the computer system. 

---