On the Proper Selection of Surrogate Measures in the Design of Data Collection Networks

Designers of data collection networks seldom can optimize the parameters of their designs on the basis of their effects on the true measures of the objectives for which data are to be collected. Therefore, surrogate measures are commonly employed to fill this vacuum. In such cases, too little concern for the communality of the surrogate and the desired effects can result in networks that not only are inefficient, but also can lead to the collection of data that have negative real impacts. Frequently, the maximization of information content is selected as a surrogate because of the rather common belief that all information is good. If this choice is made myopically, the context in which decisions are made can result in poorer decisions with greater amounts of information. Part of the context of the problem definition is the selection of which of the definitions of information content will be used -- that of minimization of statistical entropy of Shannon or that of minimization of error variance of Fisher? Examples of the shortcomings of each are used to illustrate the care that must be taken in choosing the surrogate metric of optimality.