Andre' Khuri, University of Florida
The Hazards of Coding in Linear Models
Nonsingular linear
transformations, such as coding, of the control variables in a linear
model are frequently used in regression and response surface
models. The common belief is that coding transformations do not change
the prediction capability of the model. This is not always
true. Predictions can be altered if the model is not well formulated,
which may occur if the model is reduced by a variable selection
technique. A measure will be introduced to quantify the departure of
the model from being well formulated. This measure can also be used to
provide some diagnostics that pertain to the change in the model's
prediction capability as a result of coding. (Note: This seminar
should be of interest to students in regression and linear models
courses.)