Regression hypotheses admitted to statistical evaluation should, of course, be germain to relations between the variables considered. This constraint is easily violated with the screening of arbitrarily adopted sets of hypotheses. Sequential polynomials in additive models, for example, may include components that are not commensurate with expectation (prior knowledge) and may not include those that are (e.g., unique interaction terms) and, since the analytical results from the screening of meaningless hypotheses may themselves be meaningless, there is substantial incentive for serious preanalysis consideration of the hypotheses to be evaluated.
United States Department of Agriculture, Forest Service, "Development of Structured Regression Hypotheses Interactive/Descriptive Geometry Through Five Dimensions" (1984). Miscellaney. Paper 1.