Sampling

SAMPLING

Before any experiments are conducted, the researcher must be aware of the limitations of sampling. The usefulness of any analytical method is based on the adequacy of sampling from the original population. Sampling techniques range in complexity from random methods through stratified sampling to spatial and adaptive sampling (Thompson, 2002). Sampling can be invasive, for example, thieves to remove samples from batches of powder blend, or non-invasive, as exemplified by laser optical techniques for particle sizing or spec-troscopy, which are limited by the viewing volume usually dictated by the dimensions of the laser. The bias introduced by unrepresentative sampling can be sufficient to impair decisions and lead to erroneous conclusions about a process. Consequently, representative sampling is a prerequisite to process analysis.

The use of statistics in quality control is not novel. Indeed the principles were established 70 years ago (Deming, 1938; Shewhart, 1939). These methods have since been incorporated into concepts of statistical process control (Oakland and Followell, 1986).

The basic principles of statistical analysis are beyond the scope of this volume and are the subject of a large number of foundational texts. In the realms of experimental design, Box, Hunter, and Hunter published the seminal text on “Statistics for Experimenters” in 1978. This remains a readable and informative text for those beginning to develop statistical tools to investigate processes with numerous variables.

Statistical methods mitigate the experimental difficulties associated with error (noise), confusion of correlation and causation, and complexity of the effects studied. There are many sources of experimental error that can be overcome with adequate experimental design and analysis. Frequently, exam-ples of apparent correlations occur when two variable exhibit similar patterns that may exist because of their independent relationship to a third variable. Sound principles of experimental design, specifically randomization, provide a sound basis for deducing causation. Effects are sometime so complex that they do not conform to linearity or additive interpretation. Certain experimental designs allow for interactive and nonlinear effects to be estimated with little transmission of experimental error.