Statistical Essentials—Part 4: Regression and Design of Experiments - A well-designed experiment can make it easier to understand the sources of variation. - BioPharm International
Statistical Essentials—Part 4: Regression and Design of Experiments
A well-designed experiment can make it easier to understand the sources of variation.
 Nov 1, 2008 BioPharm International Volume 21, Issue 11

DESIGN OF EXPERIMENTS

 Figure 4
Well designed experiments can reduce the risk of coming to an incorrect conclusion during a process characterization, assay validation, or process validation study. The primary goal is usually to extract the maximum amount of information regarding the factors from as few observations as possible. Typically, design of experiments can be categorized into two classes: screening designs and optimization designs. Screening designs are smaller sets of experiments that are intended to identify the critical few factors from the many potential trivial factors. A screening design assumes a linear effect, usually at two different levels or settings of the factor. Typical screening designs are called fractional factorial or Plackett-Burman designs. Optimization designs, sometimes called response surface designs, are larger experiments that investigate interactions of terms and nonlinear responses, and are conducted at more than two levels for each factor. Typical optimization designs are called central composite or Box-Behnken.

The data from a well designed experiment can be used to model the response as a function of the different factors. Regression methods as discussed previously can be applied to these data. Analysis of variance (ANOVA) is the statistical test used to assess differences in factor levels. The basis for ANOVA is the variability between factor levels compared with the average variability within a factor level.

SUMMARY

A well designed experiment makes it easier to understand different sources of variation. Analysis techniques such as regression and ANOVA help to partition the variation for predicting the response or determining if the differences seen between factor levels are more than expected when compared to the variability seen within a factor level. This paper just scratches the surface of the various statistical techniques available to the researcher or scientist. A more comprehensive course in statistics will help clarify the differences in the methods.

REFERENCES

1. Anscombe FJ. Graphs in Statistical Analysis. American Statistician, 27, 17–21.

Steven Walfish is the president of Statistical Outsourcing Services, Olney, MD, 301.325.3129,