Evaluating the Effect of Incorrect Specification of a Regression Model

Abstract

Designed experiments are frequently used in industry to identify those process variables that most strongly affect product quantity and quality. Hahn, Meeker, and Feder [3] demonstrated that it is possible to learn many things about the adequacy of a proposed experimental plan before any data are collected.

One important part of designing an experiment is to specify a statistical model to relate the observed response (e.g., yield of acceptable product) to the design variables (e.g., temperature, pressure, amount of catalyst). The data from the experiment are then used to estimate the coefficients of the assumed model and to estimate mean responses by least squares regression analysis. In the previous paper, it was assumed that the statistical model is specified correctly. In this two-part paper, evaluation of the consequences of making incorrect assumptions about the form of the model is described and illustrated with the help of a new computer program, EXPLOR II. The use of the methods in evaluating fractional factorial design alias structures is also discussed.

• Bias Evaluation
• Experimental Design
• Misspecification of Model
• Regression Analysis

Additional information

Notes on contributors

Gerald J. Hahn

Dr. Hahn is Manager, Statistics Program, Corporate Research and Development, General Electric Company. He is a fellow of ASQC.

Paul I. Feder

Dr. Feder is an associate professor in the Statistics Department of the Ohio State University. He is a former employee of General Electric Corporate Research and Development. He is a member of ASQC.

William Q. Meeker

Dr. Meeker is an assistant professor in the Statistics Department of the Iowa State University. He was formerly associated with the Industrial Administration Institute of Union College and General Electric Corporate Research and Development.